Home > AMLogic, Hardware, Linux, Testing > Amlogic S905, S905X, and S912 Processors Appear to be Limited to 1.5 GHz, not 2 GHz as Advertised

Amlogic S905, S905X, and S912 Processors Appear to be Limited to 1.5 GHz, not 2 GHz as Advertised

When I compared the performance Amlogic S905 and Rockchip RK3368 processors  last year, I noticed Amlogic S905 single thread performance not being much faster than the one of Rockchip RK3368, despite the former allegedly clocked at 2.0 GHz against the latter 1.2 GHz. But early this month one member (koschi) on ODROID forums started a thread entitled “No performance difference between 1.5, 1.75 & 2GHz” about ODROID-C2 board.

Amlogic_S905_1.5-GHz

The first post link to Ant-computing website about choosing a processor for a build farm, and one of the tables comparing various ARM and x86 processors with the claim that Amlogic S905 was listed to 1.536 GHz, because “the device claims to run at 2.016 GHz but the kernel silently ignores frequencies above 1.536 GHz!”

ARM_x86_CPU_Comparison_TableSo koschi did his own tests with 7z and sysbench, using cpufreq-set -u $FREQ to set a fixed frequency between test, and could confirm the issue:


So it looks like Amlogic is cheating with the frequencies, and anything above 1.536 GHz does not change the results… There has been many comments in the thread with others confirming the issue. Hardkernel investigated the issue, and the answer while truthful is disappointing:

Please don’t waste your valuable time.
All the Coretex-A53 based Amlogic SoCs have the same issue. S905, S905 Rev-C(S905H), S905X and even S912.
The kernel clock frequency 1.75 and 2Ghz do not exist in the BL3x blobs.
It must be Amlogic’s fault. But we should detect it earlier in our internal development stage. Really sorry about that mistake.

We’ve been modifying/testing the BL3x code to find a maximum feasible/stable clocks like 1.53Ghz, 1.58Ghz, 1.61Ghz, 1.65Ghz, 1.68Ghz, 1.71Ghz, etc for a couple of weeks.
But 1.65Ghz seems to be the maximum one due to some power and heat issues. We will run the burnA53 on Ubuntu and the Stress app on Android with 10 set of C2 boards in this weekend (for 48 hours stability test) to make sure it.
Once we will have a test result, we will report it on this thread with update package for further sampling test in early next week.

If we disable two cores, the SoC can run up to 1.75Ghz probably. But we need further stability test too.
And yes. I fully agree most people don’t like this approach.

Anyway, we know well 1.65Ghz or 1.75Ghz is still very far from the 2Ghz. So we will change the C2 specification in our home page next week as per the test result.
And we must consider some compensation for C2 users. But we need to check what we can do first.

Please accept my sincere apologies for the mistake and kindly understand our situation.

The only good news is that the limit might not be 1.5 GHz, and possibly closer to 1.7 GHz, with the final number pending Hardkernel burn-in test results. That issue will not only affect ODROID boards, but also any other devices based on Amlogic S905, S905X, and S912 processors.

Thanks to Tim for the tip.

  1. Christian
    August 28th, 2016 at 13:23 | #1

    On one side I am always amazed how cheap Allwinner and Amlogic offer their SoCs if you see what they offer especially regarding accelerated video performance comparing to what you have to pay for that in the x86 camp (often there you also have gaming performance, but I am not interested in that at all).

    On the other side they don’t play fair. All those BLOBs, license violations, lacking support for mainline kernels and now outright cheating …

  2. Ned Scott
    August 28th, 2016 at 14:45 | #2

    @Christian

    While Amlogic has done something very shady/bad here, I still wouldn’t consider them as bad as Allwinner. That’s two very different levels.

  3. theguyuk
    August 28th, 2016 at 14:47 | #3

    This has knock on effects if proven true, 3Dmark on results graph shows my mini m8s s905 (firmware 03 ( from freaktab.com update ) ) running up to 2GHz at end of tests.

    So that would mean 3Dmark tests results are in doubt too?

    Here is result screen on tinypic, right of screen

    http://i68.tinypic.com/11ce9vl.png

    I captured the result screen but don’t know how to post it here.

  4. theguyuk
    August 28th, 2016 at 15:20 | #4

    And resulting 2GHz graph from 3Dmark Ice Storm Unlimited,

    http://i67.tinypic.com/dr34g0.png

  5. neomode
    August 28th, 2016 at 15:48 | #5

    This is like Rockchip with the RK3188-T.
    Nobody where aware they exist and then it turn out they only run @ 1.4 GhZ stable.

  6. August 28th, 2016 at 16:10 | #6

    @theguyuk
    The trick here is that Amlogic kernel reports 2.02 GHz, even through it’s apparently not running at that speed. That’s why it took so long to discover the issue, as people needed to run benchmarks at various frequencies to find out.

    I had never heard about BL3x blobs before, but BL31, BL32, and BL33 blobs seems to be part of ARM trusted bootloader, and I guess Hardkernel is not allowed to release the code for it.

  7. diegor
    August 28th, 2016 at 16:38 | #7

    Ned Scott :
    @Christian
    While Amlogic has done something very shady/bad here, I still wouldn’t consider them as bad as Allwinner. That’s two very different levels.

    Do you mean that there is no gpl violation in amlogic firmware? Or are you talking about something else?

    Just wondering…

  8. Andrew
    August 28th, 2016 at 16:42 | #8

    Use an efficient cooling system ( big heatsink & fan ) to make sure temperatures stay low at all times,
    then repeat the test.

    All this may be caused by an internal CPU safety feature to prevent CPU burn. Makes sense for this mechanism to kick in as we all know how POOR is factory thermal design in all those mini systems

  9. mrmaso
    August 28th, 2016 at 17:06 | #9

    Amlogic wants to ruin their reputation? Just like allwinner did? If rockchip would get their software part right the two A Company’s could dissappear….

  10. RK
    August 28th, 2016 at 17:30 | #10

    @Andrew

    My thoughts exactly. Embedded customers glue these chips to their boards and throw them into an oven post SMT while stress-testing functionality, sensor readings & performance. This isn’t something a fab would have any hopes at getting away with so they wouldn’t even try.

    A more likely scenario is that it’s a cooling issue, software issue or even, a voltage scaling issue. Hell, for all we know they have something like Intel Trubo Boost mentioned in the Chinese fine print that didn’t carry over the translation or something… Happens all the time.

  11. mdrjr
    August 28th, 2016 at 17:35 | #11

    @RK

    Nope, its not a cooling issue.

  12. theguyuk
    August 28th, 2016 at 17:38 | #12

    @mrmaso
    Rockchip for me have been a pain. I had a 3188 TV box that I sent back ( under eBay, not as advertised ) because it was really a crippled 3188T as mentioned above. Had terrible video playback problems too.

    I had a RK3066 stick PC, for over a year for browsing and Kodi, internet TV. That was stable after I update to a unofficial firmware with same Bluetooth, WiFi chip.

    My s905, mini m8s, has been and so far is stable ( I updated to 03 firmware off freaktab, but not by the suggested method ).

    The box is good but, yes claiming 2GHz when not, leaves a bad taste in the mouth and reminds of the VW exhaust scam, maker deceives customers.

  13. Andrew
    August 28th, 2016 at 17:45 | #13

    @mdrjr

    Possible. Saying again. Possible.

    Yet, it should take much more than a statement “Nope its not that”

    It takes a test conducted with a decent PSU that will be providing a stable voltage even if there is a demand for high amperes, and an infrared external measurement tool pointed at the CPU during test, to make sure cooling heatsink is within reasonable limits at all times

  14. theguyuk
    August 28th, 2016 at 17:46 | #14

    @cnxsoft
    Does that work in the other direction too, can manufactures fool 3Dmark, by running a tweaked special one off product at 2.5GHz, but report 2GHz to 3Dmark test, to bump their product up the scores?

    Who knows?

  15. mo123
    August 28th, 2016 at 18:21 | #15

    @theguyuk
    Yes
    Here is how Allwinner cheated Antutu Benchmarks
    http://www.cnx-software.com/2015/01/11/so-this-is-how-allwinner-a80-cheats-at-antutu-benchmark/
    A user there also mentioned Amlogic has background services that detect Antutu or 3DMark and overclock the device for a while to bump up the score while the device can’t run for more than a few minutes at that speed normally.

  16. tkaiser
    August 28th, 2016 at 18:30 | #16

    @cnxsoft
    Same with RPi 3, kernel reports running at 1200 MHz while SoC in reality already throttled down to a way lower frequency. There you can never trust what’s reported through userspace utilities or sysfs (/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_cur_freq) but you have to call

    instead which reports the real cpufreq the ARM cores are running (the whole thing happens in ‘firmware’ — in other words the VideoCore does cpufreq and dvfs scaling and the kernel on the ARM cores is pretty much clueless about all of this).

    Regarding S905/S912: I don’t understand why users believe in miracles. Running 4 CPU cores at top speed generates insane amounts of heat and next to the CPU cores are video encoding/decoding engines and GPU cores all adding to the problem. So maybe Amlogic started with a budget cooling strategy allowing single-threaded CPU workloads run at up to 2 GHz while limiting multi-threaded stuff to lower cpufreqs. And then got lost somewhere due to too much problems while marketing chitchat already had been spread?

  17. RK
    August 28th, 2016 at 18:32 | #17

    @Andrew And don’t forget it could be a specific bad batch. Samsung & Motorola had a few notorious examples of those back in the day. But they happen all the time and you just don’t hear about it since the OEM calls up the sales rep, tells them they think it’s a bad batch, they hand him over to an engineer which comes over on the next airline flight and inspects the test facilities and collects samples for conformation.

    There’s procedures to these things… Not sure what gives people the impression a factory just orders 40-100k chips at 10-20$ a unit without any checks & practices in place. Sure some startup might pick up a few samples and hack on them with inexperienced engineers and no facilities. But once it’s on the fab’s floor, you have line engineers and PMIC designers that have numbers for plants world-wide since they substitute anything from resistors and capacitors to micro-controllers all the time. It’s VERY hard to get away with so much as a 3% performance drop… Let alone going form 2Ghz to 1.6Ghz…

    Very strange.

  18. August 28th, 2016 at 18:38 | #18

    @Andrew
    @mdrjr is working for Hardkernel. I think he’s well aware of the issues.

  19. theguyuk
    August 28th, 2016 at 18:42 | #19

    ARM DVFS = We just lie to customers by having the software report wrong facts.

    I am waiting for next years ARM 900099 GHz DVFS soc. 😈😉😈😉😤

  20. Marius
    August 28th, 2016 at 18:53 | #20

    From my X86 experience every chip has a sort of a ceiling where it needs a higher core voltage to go faster. The problem is even with a higher core voltage there’s a certain point where you just don’t get any further and sometimes for 100 MHz more in clock speed you have to increase the core voltage to really high levels where it’s probably not worth it.
    More voltage and better cooling will only take you so far but of course with water or liquid nitrogen you could probably get 2 GHz on this 🙂

  21. xsi
    August 28th, 2016 at 19:01 | #21

    Hardkernels response seems to indicate they knew about the issue but kept quiet aboiut it until now, when people questioned it…Why ? Were they worried about the potential impact on sales of the C2

  22. blu
    August 28th, 2016 at 19:12 | #22

    That reminds me of the RK3368 geekbox where the product is advertised as 1.5GHz, but that happens only for the big cluster, and only when the LITTLE cluster is limited to 1.1GHz, which is not what happens when you set cpufreq governor to performance – then LITTLE cluster tops at 1.2GHz and big at 1.296GHz.

  23. August 28th, 2016 at 19:28 | #23

    @Andrew / All: It’s not a cooling issue. The ODroid C2 has a proper heat sink on the S905. Additionally, Hardkernel (ODroid’s manufacturer) has confirmed that the boot blobs are capping the clock at 1.5ghz and just ignoring any values above that. Hardkernel is being rather polite about this publicly but my guess is that they’re pretty pissed and Amlogic just lost a vendor unless they can come up with a really good explanation about what’s going on, and I can’t imagine what that would be. I won’t complain if Hardkernel goes back to Exynos based ODroids, a 7870 or 8890 based board would be very niiycce (or even a 7920 for that matter).

  24. August 28th, 2016 at 19:34 | #24

    @theguyuk – Don’t conflate ARM with vendors implementing questionable ARM chips. There are good vendors out there (Samsung, Apple, Nvidia, etc), the problem is they charge more for their chips which prices them out of this market. The new Cavium based servers are a good example, they’re very fast for most applications but they also cost a few grand (http://www.servethehome.com/gigabyte-r270-t61-2u-96-core-cavium-thunderx-server-review/)

  25. blu
    August 28th, 2016 at 20:45 | #25

    @theguyuk
    DVFS has nothing to do with this.

  26. boudyka
    August 28th, 2016 at 21:46 | #26

    isn’t this always the case……with everything….dodgy cache chips (Xray’ed empty), Fake Caps with amazing uF, Bios tricked to report stuff that doesn’t exist or speeds that were never possible, DTS audio without the chip being capable and hard wired 2 channel, 4K/with all sorts of codecs speced but audio only, HD Ready, UHD Ready, HDCP compliant that just isn’t. For the British and buying TV’s that are Iplayer ready that are either not or die before the next update. Co2 emissions, say no more.

    If you believe the written spec, you are a sucker waiting to be caught, you could even say without a doubt, the written spec, is just another phishing scam….you bought it so you are just been conned!….Never believe the hype, buyer beware….there’s sharks everywhere. if its too good to be true, it probably is.

  27. theguyuk
    August 28th, 2016 at 21:48 | #27

    @blu
    Are AMLogic acting within ARM licence rules?

    Do ARM have any clause banning venders misleading customers?

    Allwinner H3 over heats does not do the original suggested 1.6 GHZ

    AMLogic s905 now also seems not to do the claimed speed.

    DVFS chip shut down cores when to hot, yet chips are sold at claimed higher speeds than can be run for long periods. ARM design!

  28. Tim
    August 28th, 2016 at 21:52 | #28

    I was curious whether the old generation Cortex-A9 boxes from Amlogic had the same “problem”. Tested a S812 box, which is sold as “2.0 GHz”. The kernel says it has these frequencies:

    96000 192000 312000 408000 504000 600000 696000 816000 912000 1008000 1104000 1200000 1296000 1416000 1512000 1608000 1800000 1992000

    From 1.6 GHz to 1.99 GHz there’s only a 6% speedup (but should be 24%). 1.8 GHz benchmarks much slower than both (why?). The other frequencies behave pretty linear compared to 1.6 GHz.

    So the actual frequency is really only 1.7 GHz and not the advertised 2.0 GHz.

    I ran the tests for a very short time, then went idle at the lowest frequency for a minute. I got repeatable results.

    Cooling is OK on this box. The PMU is a Ricoh RN5T618M. The DVFS settings in the DTB look OK.

  29. August 28th, 2016 at 21:55 | #29

    @theguyuk
    DVFS has nothing to do with this. This is a hard set cap that’s kicking in at startup. It doesn’t matter how many threads your running or what type of heatsink is in place. Your question about ARM licensing is a good one, even though this is not Arm’s fault, if enough vendors do this is makes ARM over all look bad.

  30. theguyuk
    August 28th, 2016 at 22:07 | #30

    @thesandbender
    I hear what you say but Digital Voltage Frequency Scaling was originally designed to save battery power by lowering power used. ARM and Licence holders are selling Soc chips rated at high GHz when in truth they only run at such speeds for micro seconds, longer periods and heat causes GHz to lower by design. Hence I allege they conspire by design.

  31. nobitakun
    August 28th, 2016 at 22:34 | #31

    And all of this is caused because the damn companies don’t want to move to 14nm. It is nonsense to make chips in 2016 with 28nm, a tecnology from 5 years ago. The reason is to save 5$ per chip? nonsense?

  32. August 28th, 2016 at 22:38 | #32

    @xsi
    No, Hardkernel’s response indicates that they are investigating and trying to determine exactly what’s happening. There’s a big difference between saying 7-zip see’s no performance increase between 1.5ghz and 2.0ghz and having specific confirmation that the clock speed is locked at 1.5ghz (not to belittle the person who originally found this, good investigative work). There could be some reason we haven’t anticipated, HK is just trying to figure out exactly what’s going on so they can fix it or have enough proof to go back to Amlogic and get a refund/recall/etc. You’ll note that HK is already talking about a possible refund. HK has a solid and well earned reputation in the community but it will take a while to get this all squared away properly. And I should note that I have no affiliation with HK other than I own a few ODroid boards.

  33. mlinuxguy
    August 29th, 2016 at 01:37 | #33

    The key issue with the C2 is that there are no entries for frequencies higher than 1.5ghz in the binary blob from Amlogic
    On a previous issue with lockups they had to work with Amlogic to change DVFS voltage entries and fix the buggy PWM driver (controls the CPU voltage). At the time I assumed Amlogic let them see the volt/freq tables and adjust those for the thermal issue. Having missing entries for freq > 1.5ghz suggests otherwise, that instead Amlogic did the tables for them and HK fixed the PWM driver.

    Hiding frequency control in signed binary blobs is a bad idea, it lets chip vendors get away with inflated specs. I suppose a TV set-top box re-seller doesn’t care, it’s just a marketing bullet point.

    I am campaigning for volt/freq table control from the C2’s boot.ini similar to what the RPI allows in config.txt to let the individual user choose their max clock.

    Some of us have pools of liquid nitrogen sitting around unused…. :p

  34. blu
    August 29th, 2016 at 01:55 | #34

    @theguyuk
    Conspire by design? As proven by DVFS? You’re jesting, right?

  35. GanjaBear
    August 29th, 2016 at 04:34 | #35

    Odroid/HK is a scam, C1’s USB problems have never been fixed and now this! False marketing all day long with some bait-and-switch tactics thrown in. (C1 had problems handling certain brands of SD cards so hey, they introduced C1+, all the while marketing that crap as some sort of Rpi alternative).

  36. cruz
    August 29th, 2016 at 04:51 | #36

    As a C2 owner this is disappointing. More from a hardware enthusiasts perspective. The only reason I went for the C2 instead of the Pi3 was the perceived higher performance, and I paid double, and more with customs and taxes. If it was 1.5 it would have been a closer decision.

    I don’t regret it, we can only make decisions with the information we have, but I don’t see how I can trust Amlogic moving forward and while I still trust Hardkernel I am disappointed their QA process did not get this, so this taints them slightly too.

    On a related note I think the extreme closed nature of ARM and its partners and kicking the ball back and forth, making life difficult for open source developers had damped my enthusiasm for ARM as a future potential competitor to Intel and x86.

    Many of us are expecting ARM to crash the party but their closed nature and unwillingness to divert resources or help open source developers make an ARM desktop possible can only mean there are factors beyond logic and business at play here.

    So let ARM be content being a mobile phone SOC, and work with Google and other partners in secrecy for their devices. Linux and this entire rich platform we have today happened because of open source. For hardware too we will need open source initiatives. I think struggling with ARM is a waste of effort.

  37. rasz_pl
    August 29th, 2016 at 04:54 | #37

    @boudyka

    Yes, Chinese LIE all the time. Its called Face (mianzi).
    https://www.linkedin.com/pulse/culture-matters-chinese-asians-liars-dr-vivencio-ven-ballano
    http://www.china-mike.com/chinese-culture/understanding-chinese-mind/cult-of-face/

    Someone at amlogic assumed their SoCs would be able to reach 2GHz after tapeout, this turned not to be the case, but it was already too late, marketing got hold of 2GHz number and retracting would be a HUGE blow to their image (in china). Lying is perfectly acceptable, after all who are you going to believe? some laowai idiots on on internet or successful national company?

    Chinese mAh battery ratings, gadgets spontaneously combusting because someone omitted overvoltage circuit altogether, 50W LED lights that consume 20W from the socket, all that is just another day in China.

    How can you tell if Chinese person is lying? Check if their lips are moving.

  38. August 29th, 2016 at 05:10 | #38

    @GanjaBear
    Right… because RPi’s work with every SD card out there (hint – they don’t http://elinux.org/RPi_SD_cards). SD cards aren’t just a hunk of flash memory, they have an embedded controller and sometimes there are compatibility issues. Even $3000-4000 Canon and Nikon camera’s have some cards they won’t work with and those companies have a much bigger testing budget then smaller groups like HK or even RPi have. I have few C2’s and RPi3B’s… both have their strengths and weaknesses but neither is perfect.

  39. Pepe
    August 29th, 2016 at 05:23 | #39

    GanjaBear what are you talking about? USB problems on C1? WTF? Post here forum link or visit us on IRC #odroid.
    False marketing? Are you sure? Please tell me that you’re not troll. It’s because of AMLogic. This isnt HK fault!
    Read it again!

  40. GanjaBear
    August 29th, 2016 at 05:50 | #40

    @pepe Sod off HK shill! I know exactly what I’m talking about, go read the shitty forum yourself if you need proof.

  41. Marius
    August 29th, 2016 at 05:57 | #41

    I think this is deviating from the subject here. I’m not sure what problems HK has or doesn’t have as I’ve always avoided them due to the high shipping charges.
    The main problem here is Amlogic and their lies. I remember thinking 2 GHz, who, nice, I wonder how they’re doing this when everyone else is at 1.6 max ( a fake 1.6 too from what I understand ). Now we know how Amlogic was getting those high clock speeds.
    What surprises me is that their competition didn’t expose them ( probably because they were too busy with their own cheating ? 🙂

  42. MarkW
    August 29th, 2016 at 06:18 | #42

    Why is this a surprise for anyone? When was the last time you’ve seen any Chinese product with accurate specs? It’s all gross exaggerations and bad translations. Safer to assume nothing works as advertised until proven otherwise.

  43. RK
    August 29th, 2016 at 06:47 | #43

    @cnxsoft
    With enough time on dev lists and irc channels, you’ll learn that knowing the issues doesn’t mean squat. Even experienced engineers & devs who should know better talk trash all the time. A one liner “it’s not the cooling” might be right. But take 10 minutes to go through just the irc logs of people git bisecting issues and you’ll see how fast those “it’s the x” or “it’s not the y” turn a 180.

    Personally, my money is on a simple bug in the blob. Nothing nefarious. Maybe even a bug limited to some blob releases and some batches.

    But, conspiracies make headlines better I guess :/

  44. noone
    August 29th, 2016 at 08:38 | #44

    Amlogic is an *American* company. They just do the majority of their business in China.

    I can’t wait for Risc-V to be a thing so I can put all this ARM nonsense behind me.

  45. August 29th, 2016 at 09:07 | #45

    @RK
    The problem now is that they are not trying to get a fix to 2.0 GHz, but instead only up to 1.71 GHz… This is done in collaboration with Amlogic, so it’s a bit more than just a conspiracy. The guys who tested multi-core performance reported there results in March, so it’s likely not just bad batch issue either.

  46. August 29th, 2016 at 09:12 | #46

    I also learned a few things on #odroid IRC channel. There’s no IRC log, but here’s what I remember:

    1. I did not know what BL3x was clearly, and that’s the firmware running on the ARM Cortex-M3 core inside ARMv8 processor to take care of boot and some other low levels stuffs. Hardkernel does not have the source for it, they need to ask Amlogic to modify it for them. Apparently except some parts, but they could not discuss the details publicly.

    2. They managed to “set” Amlogic S905 to 4 GHz without modifying the kernel source code. Of course a fake frequency, but the kernel had no problem returning that frequency when queried 🙂

  47. August 29th, 2016 at 10:49 | #47

    @theguyuk
    Overheating and DVFS issues are not only because of the SoC, it’s also the responsibility of the board and overall product design to cool things down. So it’s a bit different from the issue here, as even you put Amlogic S905 device in liquid nitrogen, it would still be stuck at 1.5 GHz with the current firmware.

  48. bob
    August 29th, 2016 at 11:47 | #48

    @nobitakun

    You’re complaining about AMlogic using 28nm while Broadcom is still using 40nm with the rpi3?

  49. Pepe
    August 29th, 2016 at 14:42 | #49

    @GanjaBear
    Obvious troll w/o proof. 🙂
    @Marius
    Buy it from distributor ( http://www.hardkernel.com/main/distributor.php ) some distributor like RLX provides 24 months warranty

  50. theguyuk
    August 29th, 2016 at 15:37 | #50

    @cnxsoft
    My issue with ARM DVFS and this AMLogic outing is Soc with DVFS cannot run constant all the time at their supposedly claimed higher speeds.

    Its like driving from London to Paris in a car and speeding at 200 miles a hour for 2 minutes then driving the rest of the trip at 50 miles an hour, but when you get to Paris you tell people that you drove the whole journey at 200 miles an hour.

    Now it seem AMLogic are driving the whole London journey to Paris at 50 miles an hour but claiming they did 200 miles an hour the whole journey.

    I understand what you say CNX, present software within the s905 Soc line, block real over clocking of the GHz.

    As mentioned above by Tim, can the s802, 805, 812 Socs be trusted too?

  51. JotaMG
    August 29th, 2016 at 16:16 | #51

    Marius :
    What surprises me is that their competition didn’t expose them ( probably because they were too busy with their own cheating ?

    Yes and Yes !

  52. GanjaBear
    August 29th, 2016 at 16:33 | #52

    @pepe Silence fanboy! Odroid is crap, I will never buy that shit again.

  53. August 29th, 2016 at 17:19 | #53

    @GanjaBear
    You may have had a bad experience with the board, but normally bad companies go bankrupt, and Hardkernel seem to be doing fine (although obviously I don’t know their financials, but they’ve been around for several years).

    ODROID-C2 was also the second most preferred boards, right behind Raspberry Pi 3 in a recent survey -> http://hackerboards.com/raspberry-pi-3-takes-the-cake-in-2016-hacker-sbc-survey/

    So they must be doing something right…

  54. Benjamin
    August 29th, 2016 at 18:06 | #54

    This really doesn’t surprise me. I was always skeptical about 2Ghz A53 core (especially for a lowbudget SoC like S905). That coupled with no heatsink whatsoever on most designs.

    Looks like my suspicion has been confirmed after all.

  55. Shimon
    August 29th, 2016 at 18:48 | #55

    I must agree with the Odroid bashing, I used to own a C1 but it was unstable, with many issues, the most egregious of which was freezing on mere usb keyboard disconnect or other usb devices getting dropped on connect/reconnect.

    @cnx I don’t believe the company behind odroid is doing that well or they would’ve hired a competent kernel hacker a long time ago. The announcment is mere posturing, need I remind you about that official S905x frequency spec? (1.5ghz, so it was never a secret)

  56. mdrjr
    August 29th, 2016 at 18:59 | #56

    Shimon :
    I must agree with the Odroid bashing, I used to own a C1 but it was unstable, with many issues, the most egregious of which was freezing on mere usb keyboard disconnect or other usb devices getting dropped on connect/reconnect.
    @cnx I don’t believe the company behind odroid is doing that well or they would’ve hired a competent kernel hacker a long time ago. The announcment is mere posturing, need I remind you about that official S905x frequency spec? (1.5ghz, so it was never a secret)

    Thank you for calling me incompetent 🙂 I could go for a few lines here.. but I guess it doesn’t worth explain to some people.

  57. August 29th, 2016 at 19:15 | #57

    @mdrjr fique tranquilo. Unfortunately some people will never be happy.

  58. JotaMG
    August 29th, 2016 at 19:28 | #58

    @thesandbender
    @mdrjr

    Hardkernel is responsible for what it sells, independent if it is others fault or not, ok?
    So, in this case people have EVERY RIGHT to complain and blame Hardkernel.
    You just have to be humble, apologize and compensate the one’s who feel deceived.

  59. mdrjr
    August 29th, 2016 at 19:35 | #59

    JotaMG :
    @thesandbender
    @mdrjr
    Hardkernel is responsible for what it sells, independent if it is others fault or not, ok?
    So, in this case people have EVERY RIGHT to complain and blame Hardkernel.
    You just have to be humble, apologize and compensate the one’s who feel deceived.

    Would you mind reading the official answer from us about this on the forums? (Linked in the main article)

    Not only we were the only ones making this public while everybody is quiet but there’s also the information that we are working on a compensation process for it.

    One thing is being mad, I get it.. the other is start attacking people with frivolous claims.

    • Tim
      August 29th, 2016 at 21:12 | #60

      @mdrjr

      Don’t feed the trolls mate, you do a great job.

  60. theguyuk
    August 29th, 2016 at 20:17 | #61

    I am reminded of the old Windows PC days where the only reliable CPU and combined GPU test that the average public could rely on was, could it play the latest, highest, version of DOOM, at a playable, acceptable speed.

    Not scientific, and hardware venders still cheated, but it was the main test for game playing, and the average home use, the public had.

    We need a Doom like test, Soc makers cannot cheat!

    We are all flamed, angry, bad taste in mouth, feeling deceived I suggest.

  61. cortex-a72
    August 29th, 2016 at 21:22 | #62

    Amlogic “forgot” to say that 2.0 GHz isn’t the option. Big deal. xD
    Frequency should be chosen in Non-secure world (arm recommends that), so no secure firmware components should lock the frequency. They, amlogic, should just honestly document and advertise appropriately what frequencies their SoC CPU does support.
    And for those blaming Hardkernel, I, as not having any connection to it, and looking at your whining, want to say the only thing – you are so dumb assholes, believe me, noone wants to know how stupid and neurotic you are, noone needs to look at at your annoying hysterics.

  62. tkaiser
    August 29th, 2016 at 21:52 | #63

    theguyuk :
    We are all flamed, angry, bad taste in mouth, feeling deceived I suggest.

    LOL, not at all, I ordered an hour ago an ODROID-C2 since it’s simply a great device since the Hardkernel guys care also about the important stuff (eg providing ultra-fast eMMC modules/modes — our use case is a database mirror and C2’s random IO performance is magnitudes faster than anything else available at this size)

    A few notorious SBC overclockers are upset now and that’s understandable since they do not care about real performance but only tweaking frequencies. I prefer tweaking compiler switches, settings and optimizing code since this also greatly improves performance but also the performance per watt ratio unlike overclocking. The higher the clockspeeds the lower the efficiency: http://asic-soc.blogspot.de/2008/03/process-variations-and-static-timing.html

  63. wpcprez
    August 30th, 2016 at 00:50 | #64

    @JotaMG i’m pretty sure hardkernel has owned up to any issues that amlogic brought to the table. If you made a thing (table) and the screws you bought from vendor 1 broke, you would own up to it but flaming @mdrjr and calling him incompetent is just childish. @Shimon grow up, things happen but in the grand scheme of things this isn’t a deal breaker. It’s still way faster than a rpi3 so you’re still ahead of the game

  64. NIPSZX
    August 30th, 2016 at 04:58 | #65

    I have a s812 and all I do is run Kodi at 720 max resolution. I am waiting for the s1712 in 2021 when Kodi goes to 1080

  65. Steven
    August 30th, 2016 at 12:10 | #66
  66. tkaiser
    August 30th, 2016 at 14:08 | #67

    @Steven
    Huh? It has already been confirmed that S912 is also affected and even the post title mentions S912? But who should care? Since your use case seems to be OpenSSL why not trying it with ‘work smarter not harder’? By getting Amlogic’s crypto engine to work or by choosing another SoC that supports AArch64’s crypto extensions you get not laughable 33 percent more performance as when switching from 1.5 to 2.0 GHz completely ruining the ‘performance per watt’ ratio but OpenSSL on steroids regardless how high cpufreq can be adjusted 😉

    Are aes, sha1 and sha2 listed when you query processor features? Amlogic’s crypto engine present?

    If Amlogic’s crypto engine can be used then stuff like that (or disk encryption, VPN endpoints) gets magnitudes faster, freeing up the CPU cores for other stuff and the SoC will not even get hot.

  67. theguyuk
    August 30th, 2016 at 16:22 | #68

    @tkaiser
    Come on tkaiser do you really think Soc designers and hardware venders should lie to customers and get away with it?

    Look at all the time and effort you and Armbian, helpers have had to spend getting Allwinner H3 boards to run at a sensible GHz and not over heat, as well as sorting drivers out.

    Would really have that the norm for Socs ?☺

    I still think Doom series, Quake, Unreal and Doom 3 pushed Windows PC CPU maths, GPU graphics, faster memory, bigger memory and faster Hdrives, faster DVD, better monitors, for the average public. ☺

  68. tkaiser
    August 30th, 2016 at 16:42 | #69

    @theguyuk
    Unfortunately you don’t understand the whole problem 🙂

    H3 was overheating and known for instability for only one reason: Since the only person preparing working OS images for Orange Pis in the beginning came from ODROID’s overclocker camp before. It was loboris relying on his work done for ODROIC-C1 before to start with these insane overvolting/overclocking settings that even showed lower performance when no annoying fan was used. The whole mess started with silly overclocking on H3 since Allwinner’s defaults were 1200 MHz max.

    It took the linux-sunxi community not that much time to come up with reasonable settings but the most important part was getting that this is always a trade-off. It’s 2016 and no SoC around does not suffer from overheating, so without liquid cooling or annoying fans the key to more performance is LOWERING values: VDD_CPUX core voltage. That’s important for the average user and not CPU clockspeeds since even slight increases there require huge adjustments of VDD_CPUX which increases temperatures and consumption and the only use case for this is producing benchmark numbers to compare with other geeks.

    If I want to set up a fast web server on an SBC (needs SSL, low network latency and high random IO) the route to go is to check whether crypto engines/extensions can be used on the device of choice (since it’s plain silly to let this stuff run with unoptimized code on the CPU cores anyway) and check network tunables/performance and IOPS. And then I choose an ODROID-C2, downclock the CPU cores to 500 MHz and replace a cluster made of 10 RPi 3 since the single S905 device will be faster. This might only change when Raspi’s BCM2837 can be brought up in aarch64 state and ARMv8 crypto extensions are useable there (no idea whether we will see this anytime soon), but still RPi will suffer from horribly high network latency and horribly low random IO.

    This is the stuff that matters. Not clockspeeds or overclocking 🙂

  69. Tim
    August 30th, 2016 at 17:20 | #70

    @tkaiser
    I benchmarked the DMA crypto engine on S812. It didn’t provide the expected benefit with DM-CRYPT disk encryption. It was much slower.

    The reason is DM-CRYPT can distribute the work to multiple processors. Quad-core ARM NEON AES totally crushes the (one) DMA crypto engine. That’s why Amlogic disabled this in the kernel for all SoCs after the MX.

    The DMA crypto engine is still there, even in Amlogic ARM64 SoCs, so they can claim “hardware crypto acceleration” in marketing and datasheets, but it’s useless.

    The ARM crypto extensions (AES* instructions) for the CPU itself probably cost $$ for each sold chip.

  70. tkaiser
    August 30th, 2016 at 17:51 | #71

    @Tim
    Thanks for the (discouraging) update on Amlogic’s crypto engine. Anyway: in case it’s useable with a mixed workload (again eg. an SSL enabled web server) it could still make sense to use it since it frees up the CPU cores. Benchmarking mixed workloads is not that easy but if the mix is appropriate and the workload gets somehow bottlenecked by CPU already then delegating a part of the process to the crypto engine might increase performance.

    If not then the overclocker camp has won: higher clockspeeds –> higher performance if already highly optimized software is used (if not replacing distro packages by those making use of NEON will of course be the way better tuning)

    BTW: I did some CPU benchmarking yesterday with a Samsung/Nexell Octa-Core Cortex-A53 S5P6818 (NanoPi M3). I got nice 8.227 GFLOPS with a NEON optimized Linpack bench. At just 900 MHz cpufreq (RPi 3 would score 3.6 GFLOPS at 1200 MHz with same code / optimization level). Currently I can’t use any cpufreq above 900 MHz since I would have to buy another PSU and a stronger fan to get the full 12 GFLOPS when running at the SoC’s 1400 MHz since NEON optimizations let the board deadlock already at 1000 MHz with my PSU (or maybe the board’s PMIC does an emergency power-off and nothing above 900 MHz can be used with this optimized code). It still is always a trade-off between performance and consumption/temperatures and to expect clockspeeds ‘by spec/marketing’ is most probably naive anyway.

  71. Steven
    August 30th, 2016 at 17:57 | #72

    @tkaiser
    Actually aarch64 crypto extension is under my radar for a long time.
    Amlogic S905X/S912, Rockchip RK3368, did support crypto extension and you can accelerate AES/SHA by using OpenSSL with EVP.

    The AES benchmark here is just to show that CPU frequency did not scale above 1.5GHz.

  72. Steven
    August 30th, 2016 at 18:14 | #73

    tkaiser :
    If I want to set up a fast web server on an SBC (needs SSL, low network latency and high random IO) the route to go is to check whether crypto engines/extensions can be used on the device of choice (since it’s plain silly to let this stuff run with unoptimized code on the CPU cores anyway) and check network tunables/performance and IOPS. And then I choose an ODROID-C2, downclock the CPU cores to 500 MHz and replace a cluster made of 10 RPi 3 since the single S905 device will be faster. This might only change when Raspi’s BCM2837 can be brought up in aarch64 state and ARMv8 crypto extensions are useable there (no idea whether we will see this anytime soon), but still RPi will suffer from horribly high network latency and horribly low random IO.

    Just a reminder that Amlogic S905 used by Odroid-C2 did not support armv8 crypto extension.

  73. tkaiser
    August 30th, 2016 at 18:32 | #74

    Steven :
    Just a reminder that Amlogic S905 used by Odroid-C2 did not support armv8 crypto extension.

    Sure, I was talking about RPi 3 here since some people believe they would get ARMv8 crypto extensions as soon as the cores are brought up in aarch64 state and I didn’t look into this yet at all (considering RPi 3 quite uninteresting due to lack of other features). Since you wrote only about S905X/S912 regarding crypto extensions useable with OpenSSL… does this mean no acceleration on C2/S905?

  74. Steven
    August 30th, 2016 at 18:44 | #75

    @tkaiser

    tkaiser :

    Steven :
    Just a reminder that Amlogic S905 used by Odroid-C2 did not support armv8 crypto extension.

    Sure, I was talking about RPi 3 here since some people believe they would get ARMv8 crypto extensions as soon as the cores are brought up in aarch64 state and I didn’t look into this yet at all (considering RPi 3 quite uninteresting due to lack of other features). Since you wrote only about S905X/S912 regarding crypto extensions useable with OpenSSL… does this mean no acceleration on C2/S905?

    1. armv8 crypto extensions is optional for armv8 SoC.
    2. crypto extension can be used under aarch32/aarch64 mode.
    3. S905/BCM2837 did not have armv8 crypto extensions(AES/SHA).
    4. Most ARMv8 SoC now include crypto extensions just for better looking benchmark numbers(GeekBench).

  75. August 30th, 2016 at 20:06 | #76

    @Steven And just to clarify for all reading, the S905 does support some hardware crypto acceleration, just not the ARMv8 extensions (as Steven stated). From the S905 spec sheet (Section 27.1):

    – AES block cipher with 128/192/256 Bits keys, standard 16 bytes block size and streaming ECB, CBC and CTR modes
    – DES/TDES block cipher with ECB and CBC modes supporting 64 Bits key for DES and 192 Bits key for 3DES
    – Hardware key-ladder operation and DVB-CSA for transport stream encryption
    – Built-in hardware True Random Number Generator (TRNG), CRC and SHA-1/SHA-2 (SHA-224/SHA-256) engine

  76. August 30th, 2016 at 20:25 | #77

    I just went through the whole S905 spec sheet and there are only two places where it addresses 2.0ghz (that I can find).
    Section 2 : The quad-core ARM Cortex-A53 CPU can be overdriven to 2.0GHz
    Section 3 : Quad core ARM Cortex-A53 CPU up to 2.0GHz (DVFS)
    In both cases emphasis is mine. Ominously I can’t find anything else that specifically addresses how to “overdrive” the chip or how DVFS is implemented (but I’m not an electrical engineer and have been previously banned from using a soldering iron by coworkers so take that with a grain of salt).

  77. Tim
    August 30th, 2016 at 20:32 | #78

    @thesandbender
    Yes, these are what I benchmarked. It’s kind of an extension of the DMA controller, which has a plain copy mode or a copy + encrypt + hash mode. This is only interesting for working on larger blocks or when you need to do a copy, anyway. Which is exactly the use case of DM-CRYPT.

    To clarify further, one needs to compare 4 things here:
    1. Amlogic proprietary AES/SHA via DMA engine (CONFIG_CRYPTO_AML_HW_CRYPTO)
    2. ARM crypto extensions via new instructions, separate optional IP license from ARM (CRYPTO_AES_ARM64_CE*)
    3. ARM NEON assembler-optimzed crypto code (CONFIG_CRYPTO_AES_ARM* or CRYPTO_AES_ARM64_NEON_BLK)
    4. Plain C crypto code (CONFIG_CRYPTO_AES)

    The 32 bit CPUs don’t have #2. Only the 64 bit CPUs, but it can be used in 32 bit mode user mode, too. But none of the current Amlogic S9xx SoCs have these instructions. Probably because Amlogic didn’t want to license it from ARM or saw no necessity for this in a media box (software crypto is plenty fast for that).

    Of course, #2 is the fastest, but sadly not available.

    Since the DMA engine in #1 is single-threaded, #3 on a quad-core beats it.

    #4 is much slower than #3. You don’t have to deal with this, unless the vendor forgot to properly set the kernel config option for #3. Well, the Amlogic default kernel config doesn’t set it … sigh.

    I only compared AES and not SHA1 speeds for HW/SW implementations. But that’s not very useful, except for Bitcoin mining. FPGA mining rigs are orders of magnitude ahead of any CPU … and even these have trouble making a profit anymore.

  78. Tim
    August 30th, 2016 at 21:10 | #79

    @Tim
    I forgot to mention that #1 needs CONFIG_CRYPTO_AML_HW_CRYPTO **and** a DTB entry. Otherwise it’s not activated.

    DTB for S8xx:

    aml_aes {
    compatible = “amlogic,aes”;
    dev_name = “aml_aes”;
    };

    aml_tdes {
    compatible = “amlogic,des,tdes”;
    dev_name = “aml_tdes”;
    };

    DTB for S9xx is similar, but needs “amlogic,aes_dma”, “amlogic,sha_dma” and “amlogic,des_dma,tdes_dma” in the compatible lines. I haven’t tried. In some other thread mentioned above there’s a fix for the ARM64 kernel driver, since it’s broken.

  79. blu
    August 30th, 2016 at 21:45 | #80

    Re the crypto ISA extension, it can’t be that costly an IP, given Rockchip have had it since their first ARMv8. Apparently Amlogic think whatever they have (i.e. the DMA crypto engine) is sufficient for them to not have to license anything. That’s just a guess here, of course.

  80. Winston
    August 31st, 2016 at 09:56 | #81

    It looks like the addition of the S912 to this benchmark table might be interesting:

    Amlogic S905 vs S812 Benchmarks Comparison

    http://www.cnx-software.com/2016/01/29/amlogic-s905-vs-s812-benchmarks-comparison/

    Of course, just be sure to get rid of their bogus 2GHz claim in the heading.

    Because they’re all apparently limited to the same maximum clock speed, I begin to think that there’ll be little significant difference in benchmarks between any of the S812/905/912 SOCs as long as they all have the same amount of RAM (2G seems to improve things).

    The S905 looks significantly better than the S812 in that benchmark table linked to above ONLY in 2D graphics rendering which would be useful for retrogaming emulation (MAME, NES, SNES), but do the emulators actually need it? I doubt it since I suspect most of the game playability would be more related to the efficiency of the emulation code than to 2D rendering speed.

    So, I guess it’s mostly the more unusual TV box _features_ that one should look for then, primarily gigabit ethernet and BT4, and pretty much ignore S812 vs S905 vs S912 although the latter two would tend to be the ones associated with those unusual features since they’re the newer SOCs.

  81. August 31st, 2016 at 10:02 | #82

    @Winston
    While the amount of RAM may affect overall system performance, it does not usually impact benchmarks. A system with 1GB RAM will score the same as one with 2GB RAM since only one main application (the benchmark) is running.

  82. Winston
    August 31st, 2016 at 10:07 | #83

    @cnxsoft Do 3D game emulations push the 1G RAM limit? If not, considering that at least in my case I’d be running one app at a time, would 2G be worth the additional expense?

  83. August 31st, 2016 at 10:24 | #84

    @Winston
    It’s possible. I’m not sure.
    If you don’t have enough memory in Android, the system will normally kill the app to free up memory. Some people have reported issues with Kodi being killed with just 1GB RAM, but that’s not my experience, except on M-195 TV box. It’s also possible that overtime, as you install some apps, there are more background tasks, and issues related to low memory become more frequent.

    I know some people who would not buy boxes with less than 2GB RAM.

  84. August 31st, 2016 at 10:26 | #85

    At least we’ve been noticed… One manufacturer told me:

    Amlogic update the specification for S905 S905x and S912 processors only to 1.5GHz instead of 2.0GHz, we will update all information.

  85. Winston
    August 31st, 2016 at 10:43 | #86
  86. theguyuk
    August 31st, 2016 at 13:25 | #87

    @cnxsoft
    Do the CPU and GPU share system memory?

    If GPU has own memory then fine, but if shared and as resolution of display increases the shared 1GB ram gets squeezed with the GPU needing more memory, in old Windows PC design.

    Is ARM GPU memory use the same?

  87. tkaiser
    August 31st, 2016 at 15:24 | #88

    @theguyuk
    CPU, GPU and video engines share access to DRAM (that’s the reason why memory bandwidth for CPU increases if you set framebuffer size and depth to the minimum or even disable display/GPU completely which is something that gives you even a bit more performance on headless devices/servers based on cheap tablet/OTT SoCs).

    BTW: My Odroid-C2 arrived in the meantime 🙂 Super fast eMMC (really rocks). Tried out the NEON optimized Linpack (OpenBLAS-0.2.18/hpl-2.2), 20 runs with passive cooling only: average SoC temp 48.2°C and 3.9424 GFLOPS:

    With an annoying fan combined I got constant 37°C SoC temp and only 3.949 GFLOPS. Hmm… a cheaper NanoPi M3 scores 8.227 GFLOPS at 900 MHz. If count of CPU cores (4 vs 8) and clockspeeds (1536 vs. 900 MHz) are considered it gets really weird since then a ‘1 GFLOPS per core’ would need the S905 to be clocked with 1.556 MHz while M3’s S5P6818 could be clocked with just 870 MHz while running ARMv7 optimized code (no ARMv8/64-bit kernel/userland available yet).

    There must be something wrong, maybe the NEON optimizations work better with code generated for ARMv7 (assembler)? The low temperatures would indicate the same…

  88. no_spam_for_me
    September 1st, 2016 at 17:45 | #89

    @neomode
    Not only RK3188… also the RK3288 doesn’t support the max frequency they promise (only with a hack)… anyway, all ARM SoC throttles down after a few sec because of heat… so what…

  89. blu
    September 1st, 2016 at 17:59 | #90

    @tkaiser
    What you’re observing could be as “trivial” as compiler version/control options. For the record, A64 NEON has outperformed A32 NEON on every platform I’ve tested yet. If you really want to load your C2 with densely-executed NEON code for reference purposes, drop me a line via Jean-Luc – I wrote to him once about microbenchmarks, so he has both my email address and the relevant links.

  90. Winston
    September 1st, 2016 at 22:24 | #91

    no_spam_for_me :
    @neomode
    Not only RK3188… also the RK3288 doesn’t support the max frequency they promise (only with a hack)… anyway, all ARM SoC throttles down after a few sec because of heat… so what…

    I think what should come from this is widespread notice to the audience of SBC and Android TV box reviews that the claimed clock speeds are effectively bogus and that newer SOCs aren’t necessarily better and can actually be worse performers than older SOCs. Example: the S812 vs S905 benchmarks found elsewhere here.

  91. theguyuk
    September 2nd, 2016 at 15:27 | #92

    @Winston
    Very well put true points.

    I was becoming a s905 recommender by purchase experience but now, I wonder what other corners have been cut or lied about.

    Sound?

    True ROM size?

    Electrical safety?

    Security of user data in AMLogic Android, the list could go on.

    Once AMLogic lie and cheat the doors to customer abuse are forever open.

    IMHO.

  92. JotaMG
    September 2nd, 2016 at 17:04 | #93

    @theguyuk

    It is not only AMLogic, it is many more corporations (VW/Audi/etc) that lie about their specs…
    That is becoming as a normal part of life.
    The important is whether it is intentional or just incompetence, and how they react and compensate.

  93. tkaiser
    September 2nd, 2016 at 17:05 | #94

    @Winston
    The target audience of this type of SoCs is absolutely clueless what matters and buys numbers. 64-bit is ‘better’ than 32-bit since twice as much, more CPU cores are better than less since ‘much’, 2 GHz are better than 1.8 since…

    Everyone who put a finger on one of these SoCs knows that ‘octa-core’ or now ‘deca-core’ combined with 2GHz in TV boxes does not work. It’s either all cores at low clockspeed or few at full clockspeed since TV boxes are neither equipped with cooling capabilities based on liquid nitrogen nor 40W PSUs.

    We know how this works from a technical point of view and we should try to understand the sales side. A vendor producing the best TV box SoC around with 2 CPU cores at 1 GHz (enough), superiour video engine and GPU cores (the stuff that really matters) will sell exactly nothing. Because not enough numbers. Everything as expected and specs will be throttled down in reality by cooling and power requirements even more as time goes by since the average TV box customers still buys only numbers and numbers have to increase. Only accepted decrease is the price.

  94. theguyuk
    September 2nd, 2016 at 17:34 | #95

    It says a lot about how are customers to be able to know or recognise the truth when this has took so long to come out and so many technical people and benchmark software writers have been fooled.

  95. no_spam_for_me
    September 2nd, 2016 at 17:46 | #96

    Winston :

    no_spam_for_me :
    @neomode
    Not only RK3188… also the RK3288 doesn’t support the max frequency they promise (only with a hack)… anyway, all ARM SoC throttles down after a few sec because of heat… so what…

    I think what should come from this is widespread notice to the audience of SBC and Android TV box reviews that the claimed clock speeds are effectively bogus and that newer SOCs aren’t necessarily better and can actually be worse performers than older SOCs. Example: the S812 vs S905 benchmarks found elsewhere here.

    hmmm, the S812 isn’t really faster than the S905 because at the CPU part (and because of clock speed we talk about the CPU part at the SoC) they have nearly the same performance… at GPU the S812 is better at 3D games than at the S905… (I know the S812 vs S905 benchmarks found here http://www.cnx-software.com/2016/01/29/amlogic-s905-vs-s812-benchmarks-comparison/)
    Anyway, clock speed isn’t all…

  96. cortex-a72
    September 2nd, 2016 at 21:39 | #97

    @JotaMG
    Of course it’s intentional. They, amlogic. wanted to run it on 2.0 GHz. They start to advertize it as 2.0 GHz. Then, they saw, it burns at such speeds. Then they put modifications in early stage blobs to silently ignore establiching anything higher than 1.65 GHz (and voltages possibly too). And still they contiinue to sell it as 2.0 GHz capable. It’s an intentional cheating and lying.
    I don’t know how DVFS is done in linux, but it should be done exactly in the non-secure OS, not in the secure blobs. as opposed to resetting. Seems linux asks these blob to do this and the latter cheats the kernel. Or, they modified the kernel too. How this can be “unintentional”?

  97. Tim
    September 2nd, 2016 at 21:43 | #98

    It’s intentional, no doubt about it.

  98. tkaiser
    September 2nd, 2016 at 23:17 | #99

    @cortex-a72
    There is no ‘DVFS in Linux’ since it depends on the SoC in question. On the recent Amlogic SoCs a Cortex-M is doing the job just like on RPi 3 (there the primary OS called ‘firmware’ running on the VideoCore IV does everything, the ARM cores also have not the slightest idea what’s going on and the kernel also cheats on you reporting running at 1200 MHz while it’s already at 900 MHz due to throttling!), on Allwinner SoCs it’s different and the kernel controls behaviour: SoCs for mobile usage like A64 have an accompanied PMIC that does voltage regulation (CPU talks to it via an I2C variant), with SoCs for OTT boxes it depends: More than 1 variant possible, the device vendor can choose between doing it crappy (no DVFS at all), superiour (adjusting voltage in 20mV steps) and something in between (a few steps). In any case the kernel has full control here and it’s the developers responsibility to choose dvfs operating points that are balanced between reliability (voltage high enough) and performance (voltage as low as possible — every mV more insanely increases consumption and temperatures)

    But this doesn’t change a bit regarding the real problem since its 2016, all modern SoCs overheat like hell and these high clockspeeds require VDD_CPUX adjustments that make them absolutely inefficient. So the advertised clockspeeds are BS anyway from a technical point of view. Especially considering the real use case for these SoCs: Android TV boxes where nearly everything that matters happens in different SoC parts. High clockspeeds might matter for Antutu scores but are more or less irrelevant for normal scenarios since no one wants to buy TV boxes with annoying fans.

    It’s only the overclocker’s camp that is disappointed now (and it’s kinda funny it took them so long to realize this). For the average TV box customer this whole CPU MHz thing is of no relevance at all 🙂

  99. Marius
    September 2nd, 2016 at 23:43 | #100

    @tkaiser
    By following this (AM)logic they should have stated it runs at 3 GHz and then run it at half because well it doesn’t matter for the users of TV boxes anyway, right ?
    People do buy specs because they expect better spec mean better performance and when you see the same A53 but at 2 GHz as opposed to 1.5 GHz you go … I’ll get this one.
    OK there might be a thing with overheating and stuff so even if it’s 2 GHz capable it will run at 1.5 GHz if the manufacturer doesn’t provide proper cooling. This is then the responsibility of TV box or tablet manufacturer and they’re the ones lying that it’s running at 2 GHz.
    I do expect however that if I slap on a big heatsink or a one with a fan it will run at 2 GHz.

    Also let’s not forget that these are TV boxes. You can make an aluminium enclosing for them and connect the SoC to the case with a heatpipe or some thermal pads.

    I’ve seen mini PCs with Core i5 ( low power ones ) being passively cooled but their aluminium enclosing so surely this can handle some quad core A53s ….

    Finally I’d just like to add that Mediatek made the Helio X10 with 8 A53 cores that can run at up to 2.2 GHz ( made in 28 nm also ) . Now they might not all run at the frequency for extended periods of time but it’s certainly doable so Amlogic has no technical excuses for not being able to run their latest quad core at 2 GHz more than a year after the X10 made its debut.

  100. tkaiser
    September 3rd, 2016 at 00:56 | #101

    @Marius
    Remember those times in x86 land? When Pentium IV clocked at 4.0 GHz and you needed nothing else to heat up your home in Winter than 1 or maybe 2 PC? And then this small company Intel bought came up with x86 designs that were faster at a fraction of clockspeeds? And even Intel realized they had to shut down their radiator business to focus on performance instead?

    What’s this clockspeed hype all about? Yeah, now we repeat the same mania with ARM. And it’s as silly as back then to focus on clockspeeds instead of real performance. One Soc outperforms the other running at lower clockspeeds. So why not looking at real performance (when needed) and performance per watt ratio (when interesting)? And forget about clockspeeds at all?

    If a SoC vendor in this specific segment now wants to outperform another in terms of sales and not performance then he has to go the Pentium IV route since both the real target audience (TV box customers) and the irrelevant few SBC people buy only numbers (just like back then with PCs). It’s not about real performance at all. I did yesterday some database benchmarks on the slowest eMMC module available for ODROID-C2. AWESOME! CPU clockspeed my a**! 😉

  101. tkaiser
    September 3rd, 2016 at 02:03 | #102

    @Marius
    Ah, almost forgot: ‘TV box, aluminium enclosure, heatpipe, thermal pads’… Why on earth? We’re still talking about SoCs made for a specific purpose, aren’t we? 🙂

    The CPU cores in a TV box SoC are the most irrelevant parts. They have to be there and their specs have to look beefy since the target audience knows nothing. Take a rather expensive RPi 3 and an $8 Nano Pi NEO and start the TV box contest decoding HEVC. The old and incapable SoC (RPi’s BCM2837 which is the the same as BCM2836 which is the same old boring BCM2835 just with exchanged ARM cores) needs overclocking, heatsink and fan to do this (maxing out CPU and GPU).

    NanoPi’s newer SoC (pretty boring and slow BTW) does the same job on the SoC’s video decoding engine with a slight increase of temperatures and consumption while idling aroung when looking at CPU utilization. We’re still talking about SoCs for a specific purpose and the CPU cores in the SoC are pretty much irrelevant here (they’re slow ARM cores anyway!)

  102. Winston
    September 3rd, 2016 at 03:16 | #103

    “For the average TV box customer this whole CPU MHz thing is of no relevance at all.”

    I agree with everything you said except that. In a way, it does make a difference when they choose to pay more for a new box with a “new and improved” SOC that isn’t necessarily better at a claimed higher clock speed that’s bogus.

    However, this marketing issue certainly isn’t unique to Android TV boxes. There’s the pixel race in cameras, a bazillion pixels in a puny sensor, noisy pixels though requiring all kinds of digital noise reduction which can then lead to all kinds of bizarre artifacts and detail loss, grass blades and gravel becoming very weirdly smoooooth, but just in patches.

  103. September 3rd, 2016 at 09:40 | #104

    @tkaiser
    I agree that CPU frequency and number of CPU core are overrated and not that useful in TV boxes. The leaders for the STB SoC market (Broadcom/STMicro) actually understand since their STB processors often only have two cores.

    However, it’s still not right to lie about the CPU frequency in order to sell more devices, and it’s hard to see how this could be an honest mistake.

  104. theguyuk
    September 3rd, 2016 at 17:37 | #105

    The AMLogic venders cheating customers might have another cheat developing.

    A internet search show some suppliers advertising Kodi 17 installed , as far as I know that version and above is still beta and in development.

    I joined Google Beta testing option in Google Play under a month ago, and I have Kodi beta 17, with bugs and still some function don’t work. It is still in development. As Google tell all testers!

    I only have a 2GB s905 Android 5.1.1. I cannot comment on the other socs.

    Is this yet more customer cheating, with unstable software which is not finished?

  105. theguyuk
    September 3rd, 2016 at 17:45 | #106

    @tkaiser
    So is what that translates in to, that customers just wanting browsing and video viewing, playing, might as well just buy the cheapest 2GB, stable, Android TV box they can? Even Android 4.1 – 4.4 .

    I suggest 2GB, due to I feel more memory, does help slightly with buffering in video?

  106. xxx
    September 3rd, 2016 at 19:17 | #107

    What is about S912,very bad antutu results.I think it not have 2.0 ghz.

  107. mo123
    September 3rd, 2016 at 19:48 | #108

    @xxx
    Please read the headline.
    It also only supports 1.5Ghz.

  108. roel
    September 4th, 2016 at 02:44 | #109

    tkaiser :
    @Marius
    Remember those times in x86 land? When Pentium IV clocked at 4.0 GHz and you needed nothing else to heat up your home in Winter than 1 or maybe 2 PC? And then this small company Intel bought came up with x86 designs that were faster at a fraction of clockspeeds? And even Intel realized they had to shut down their radiator business to focus on performance instead?
    What’s this clockspeed hype all about? Yeah, now we repeat the same mania with ARM. And it’s as silly as back then to focus on clockspeeds instead of real performance. One Soc outperforms the other running at lower clockspeeds. So why not looking at real performance (when needed) and performance per watt ratio (when interesting)? And forget about clockspeeds at all?
    If a SoC vendor in this specific segment now wants to outperform another in terms of sales and not performance then he has to go the Pentium IV route since both the real target audience (TV box customers) and the irrelevant few SBC people buy only numbers (just like back then with PCs). It’s not about real performance at all. I did yesterday some database benchmarks on the slowest eMMC module available for ODROID-C2. AWESOME! CPU clockspeed my a**!

    +1, Love it. Still have such a pentium 4 running (stupid of course regarding power/watt ratio), but with a SSD it doesn’t feel sluggish at all. Only it sounds like an airplane…

    In my experience I/O-performance is more important then high cpu speeds. It’s only when you do certain tasks that need heavy calculations (like CAD) You notice the difference. But I would never buy a ARM based pc to do CAD.
    If You really need the 2 Ghz of the odroid, then I think you are shopping in the wrong part of the shop.
    It’s true they lied, as most companys do and to be honest we all know. How much woman are there who are as beautiful without as with make-up? You only see it the morning after…

  109. theguyuk
    September 4th, 2016 at 15:10 | #110

    There is a 95 min video with Steve Furber, he is one of the people who brought use RISC ARM CPU. He comments the important part is how fast the CPU can get instructs to and from memory and how little time it take the CPU to do instructions. With present day CPU in built memory cashe helps too.

    If you are interested here is a link to the video, http://armdevices.net/2016/08/22/arm-architect-steve-furber/

  110. Theguyuk
    September 5th, 2016 at 18:32 | #111

    @cnx-software

    Hello ☺

    Well I have a Remix Os mini here at my home, purchased off Gearbest.

    I ran 3Dmark and given the Remix Mini is a Allwinner 4 x A53 at 1.2GHz and only a Mali400 MP2 scores are a lot lower than a cheating s905 AMLogic.

    I have updated Remix Os to latest version based on Android 5.1.1

    I desided out of random curiosity to download free version of Riptide GP2 off Amazon underground and found the Remix mini can run at all the highest settings my mini m8s s905 can. The Remix mini top gets slightly warm to touch but game plays fine despite Remini mini having a lower GPU and CPU mix.

    So can you or anyone recommend a free game which really challenges the s905 claimed Mali500 penta cores. So I can compare running it on a Remix Mini, given that S905 cheats benchmark tests.

    I am curious to try a comparison!

  111. Theguyuk
    September 5th, 2016 at 19:12 | #112

    @Theguyuk
    Should read Mali 450 claimed Penta cores.

  112. tkaiser
    September 5th, 2016 at 19:24 | #113

    @Theguyuk
    Great that you start to look into the right direction 😉

    Your Remix Mini is said to run at 1.2 GHz but THS/dvfs/cpufreq default settings limit the CPU cores to 1152 MHz max (so it’s time to start the next cheating campaign, right?)

    Since we all (at least in the meantime should) know CPU cores are not relevant for TV boxes from a technical point of view (it’s only about sales, clueless people demand cores and GHz). You might be surprised but even 8 year old slow and boring Mali400MP2 is able to produce insane amounts of heat. When we did DRAM reliability testing for some H3 devices (equipped with the same Mali) in the past we let the Mali400 run at 600 MHz and executed some demanding stuff on the 2 GPU cores. Temperatures increased even more than when running cpuburn on all 4 CPU cores and Allwinner’s budget cooling strategy throttled CPU cores down to 240 MHz (only way to deal with that since we were mean and forbid the kernel to downclock GPU cores which would’ve been the correct strategy).

    This is on a slow passively cooled H3 TV box (2 GPU cores running at 600 MHz with demanding stuff, SoC temperature ~95°C, 4 CPU cores downclocked to 240 MHz). Now imagine you repeat this test on faster but still slow H64 SoC as in your Remix Mini. And then try to imagine how things might look on a really fast S912 box. In a living room and not in a lab trying to fight the heat with liquid nitrogen or monster heatsinks and a battery of fans around.

    All these clockspeed specs and the race for more cores are outright BS. None of these devices is able to run even moderate loads on all CPU and GPU cores in parallel without massive downclocking or shutting cores down. They’re only able to operate as desired when the tasks in question are moved away from CPU cores (GPU and video encoding/decoding engine… and even then throttling might occur). But since the average customer is clueless all OTT box SoC vendors are forced to develop hardware with ‘either or’ specs (count of cores vs. clockspeeds) since only then marketing/sales can fulfill customer demands. And no customer knows that ‘X cores running at up to Y MHz’ translates to ‘Never at the same time’.

    Technical reality: to increase the performance of an OTT box it’s necessary to lower heat emission of GPU cores and video engine. Everything else is pretty much irrelevant. And now try to create a marketing campaign based on technical reality while your competitor announces a new SoC that allows TV box vendors to sell ‘deca-core’ boxes 😉 (also see cnxsoft’s comment regarding STB SoC market above — there technical reality matters)

  113. September 5th, 2016 at 21:14 | #114

    @tkaiser I don’t know why you guys keep on saying that core speed and core count aren’t useful for TV boxes. Under ideal cases this is true but in practice it’s not. VPUs are awesome when they implement a codec and implement it completely/properly. When you get a video that has a codec that is not fully implemented or not implemented at all, you have to fall back on the CPU to do the decoding and this is when cores and clock speed count. Hardware always lags behind software so you’re not going to find many affordable VPU IP’s that support VP9, H.265, etc. The tests on this site are a clear indication of that, a lot of boxes do great until they hit a subset of video types and then they fall over. Sometimes that because of a bad driver but just as often the VPU just doesn’t support the codec.
    Also to your comment about the Mali 400 running hot. Heat is not just a factor of clock speed, it’s also related to trace size, substrate material, etc. A 200Mhz Pentium Pro II kicked out 44W, which was a lot at the time (the barn sized heat sinks we have today just weren’t around then). Part of the reason was it had a .35 micro-meter trace. That’s 10x the size of even the cheapest SOC now.

  114. Winston
    September 5th, 2016 at 21:37 | #115

    @tkaiser This sounds like a job for water cooling! 😉

  115. tkaiser
    September 5th, 2016 at 22:52 | #116

    @thesandbender
    Ah, ok. That was new to me that it’s even possible to decode VP9/H.265 in software (in a way it doesn’t suck due to artefacts, fps too low and so on). I thought if one has a S905 based TV box and wants to watch VP9 he throws the S905 box in the bin and buys one with S905X instead (even if blogs like this reported S905X would be slower). BTW: somewhat disturbing reading me questioning Amlogic SoCs exceeding the 1.5Ghz 5 weeks ago. This is cnxsoft’s blog post with benchmark results showing S905X being slower than its predecessor: http://www.cnx-software.com/2016/07/31/amlogic-s905-vs-s905x-benchmarks-comparison/#comment-529581

  116. Theguyuk
    September 6th, 2016 at 00:07 | #117

    @tkaiser
    Claiming 2GHz when it only does 1.5GHz is a much bigger cheat. ☺

    People are buying more cores and alleged better GPU because of 4K TV chasing and Android TV boxes are starting to fill the promised casual game console market Ouya promised and aimed for. Others are cable cutters, going straight to source.

    The AMLogic sham 2GHz fooled more than casual buyers, they fooled the Benchmark software and product reviewers people trust.

    If people just want video players they can buy a NowTV, Roku, Apple TV, Amazon, box.

    At the moment my interest is outside failed, cheated benchmark software, where are the complex 3D graphics games that challenge the alleged higher end GPU.

    Android has gone 4.4, 5.1.1, 6 and now 7 quickly. So are people being sold GPU that no use for exists in Android. When that uses, programme, game, comes will their box and Android version be obsolete.

    4K needs seem to be met by the lowly Mali 400, 450.

    We don’t need water cooling jokes, it is just bad design and using battery power saving DVFS to stop overheating Socs instead of proper designed heat control.

    These Android TV boxes run with less heat than old school passive cooled graphics cards did. Cooling is not the problem, it is deeper than that. The problem is bad design at source and system board level.

    IMHO

  117. September 6th, 2016 at 00:53 | #118

    I don’t understand the logic that just because Android is on version 7 older GPU’s don’t have a use. Android doesn’t drop support for them and Android 7 will still leverage a Mail 400 for GPU acceleration. It’s more than fast enough to improve the user experience for most applications (even simple things like a drop down box are hardware accelerated if possible… when you press the drop down, the content behind it is bitblt’ed off and then bitblt’ed back when you close the drop down. Generally the application doesn’t need to do a redraw. Similar things are done for scrolling, etc).
    I agree with you about cooling. That’s one of the things I like about RPi, ODroid, etc over packaged set top boxes, there is some opportunity to improve the cooling on them. Unfortunately there’s really no middle ground in the ARM world, you’re either on lowend SoC’s or your coughing up $$$ for a Cavium. I’m guessing the market for a ~$200 board with proper SATA, upgrade-able RAM, etc really isn’t there. Closest you can get is a Tegra TX1 which is $600 and doesn’t even come with a SATA port (just M.2 though you can add a controller via the PCI-Ex4).

  118. Theguyuk
    September 6th, 2016 at 01:56 | #119

    @thesandbender
    Older devices when in Google play get sometimes told this app does not support your device. 12 to 24 months is a long time in ARM hardware, hence older stuff out dates quicker.

    You can still watch TV and Kodi on a Android 4 .0 , RK3066 1GB TV dongle / stick PC. But would you recommend one over today’s Android TV device ? Even if it meets all your needs newer hardware and Android version cost the same or very little more.

    It all depends what you are wanting to do with the device.

  119. tkaiser
    September 6th, 2016 at 02:30 | #120

    thesandbender :
    I’m guessing the market for a ~$200 board with proper SATA, upgrade-able RAM, etc really isn’t there. Closest you can get is a Tegra TX1 which is $600 and doesn’t even come with a SATA port (just M.2 though you can add a controller via the PCI-Ex4).

    Upgradeable RAM not but the closest you get is based on ARMADA 38x or i.MX6 then. Solid-Run’s ClearFog Base/Pro show excellent SATA/USB3/GbE performance and you can use mechanical converters and some u-boot voodoo to use both M.2 and mPCIe slot(s) with normal SATA disks. The quad-core i.MX6 variants show way slower SATA throughput and GbE is AFAIK still limited to 400 Mbits/sec unless you use a PCIe attached additional NIC (like it’s done on the Utilite Pro for example).

    The former are headless, the latter maybe suck regarding video/GPU (really don’t know). I wonder whether devices are out there based on ARMADA 1500 since based on specs these SoCs have both excellent multimedia and network/storage capabilities.

  120. September 6th, 2016 at 02:46 | #121

    @tkaiser The ClearFog is a nice piece of kit but it’s intended for networking, it’s a dual core 32-bit A9 with 1GB of ram. Not trying to knock it, it’s a great router platform, but it’s purpose built for networking.

  121. Winston
    September 6th, 2016 at 05:46 | #122

    @Theguyuk “We don’t need water cooling jokes”

    Somewhere on YouTube, I saw a water cooled Pi3. I thought the idea of water cooling a $30-40 SBC to be so absolutely absurd that it was funny. Apparently, so did the video’s creator.

    When I mentioned this clocking fiasco to an on-line reviewer of cheap Chinese Android TV boxes, he replied, “But look at what you get for $30” and that gave me a much better perspective on the issue.

    My philosophy is that there is no use getting upset over things that simply will not change no matter how much we talk about the issue here. New and truly more capable SOCs will only appear as advancing technology makes them possible and demand makes them profitable, not based upon what we say here. Meanwhile, marketing departments will continue to prey upon ignorant consumers as they always have.

  122. Marius
    September 6th, 2016 at 05:57 | #123

    @Winston
    Forget water cooling, this chips are for TV boxes. This have a relatively large surface size and with an extruded alu case this would be more than OK. I’ve seen 15W low power Intel chips cooled passively using only an extruded Alu case.
    Of course you need good contact between the case and the soc, maybe even a sort of a heatpipe or something and this would increase costs quite a bit.
    I do believe for an extra $10-$20 you can get proper passive cooling that would allow these SOCs to run with everything at max, CPU and GPU.
    However this might not really be needed and also let’s not forget that not matter what you do to them these Amlogic chips just can’t hit 2 GHz under normal conditions ( maybe with a peltier or liquid nitrogen, lol ).
    There are 2 problems there:

    1. The maximum clock speeds often declared are purely theoretical because manufacturers don’t provide the adequate cooling needed so the chips goes into thermal throttling

    2. Amlogic lied and 2 GHz is not even the purely theoretical maximum limit.

    Out of these thow 1 is a really known problem and it’s a sort of c’est la vie but 2 is quite new and it was quite daring of them to do it.

    It’s a bit like the VW scandal … yes fuel usage cheating was a known thing and everybody was doing it so it was taken for granted … but emissions cheating was something no one was found to be doing till that date and it was considered much worse. Also with fuel usage it’s debatable because if you give it to the right driver he might get the values declared in the specs or even lower … you probably won’t get them in real life though.

    A bit like the maximum declared frequency of an ARM chips, possible in theory but highly unlikely that your device will be able to sustain it.

  123. September 6th, 2016 at 09:55 | #124

    @Theguyuk
    The frame rate in Riptide GP2 should be quite lower with ARM Mali-400MP2 compared to ARM Mali-450MP with max settings. Can you check Remix OS framebuffer resolution? If it is 1280×720, that may explain why you get good performance in the game, since all S905 devices set it to 1920×1080.

    I already find Amlogic S905 TV boxes do not run Riptide GP2 optimally at 1920×1080 resolution. If you play it on a recent phone, or Nvidia Shield Android TV, you’ll clearly see a big difference.

  124. September 6th, 2016 at 10:05 | #125

    @tkaiser
    VP9/H.265 software decode is not possible at 4K, but should work fine for 720p, and in some cases 1080p. If that happens all cores are used, but normally you’d want to avoid that.

    But there are other videos with codec seldom implemented in hardware such as 10-bit H.264 and MSMPEG4v2, where software decode is the only option most of the time, and again all cores are used. This can also work well with proper cooling however.

    Android TV boxes can also be used for other purpose like web browser, checking email, gaming…, and may benefit from multiple cores and higher frequencies. Based on some tests I did a few years ago on an octa-core CPU (Allwinner A80), games often use only two CPU cores, sometimes four CPU cores, and usually the low power ones, since most of the heavily lifting is done by the GPU.

  125. Theguyuk
    September 6th, 2016 at 15:50 | #126

    @cnxsoft
    You would have to to explain to me how to do that.

    Also I am on the latest version of Remix OS.

    I have the free Amazon underground GP2. ( unsure if updated or different to Google Play version )

  126. Theguyuk
    September 6th, 2016 at 16:04 | #127

    Dismissing cheating just means honest companies go broke.

    Before UK food standards people added chalk to flour, put candle wax in chocolate for kids. If you accept fraud you weaken the competition.
    What next? Heart tablets that are flour, boob implants that are industrial silicone, cars that have no crash safety, planes with defective engines. Fraud is fraud even if it is just stealing 1 dollar from every charity box.

    @Winston

  127. Theguyuk
    September 6th, 2016 at 16:12 | #128

    @cnxsoft
    Way back when we had Intel, AMD and Cyrix , the Cyrix CPU were cheap but terrible slow in 3D maths games Doom, Quake, Duke Nukem 3D etc. So I paired mine with a 3DFX graphics card.

    Just like you mention the card took the hard 3D graphics work off the CPU meaning the game became playable.

  128. Roel
    September 6th, 2016 at 17:19 | #129

    @Marius
    They shoud all put the SOC on the “bottom” side of the board like the banana pro and the NEO. It’s much easier to cool it that way in a alu or steel enclosure. In that case the cooler isn’t in the way for add-on boards. On my odroid C1 the cooler is touching the bottom of my witty-pi board. I should mill of the cooler to make it fit properly, didn’t do it yet as I assume it will influence the cooling.

  129. Roel
    September 6th, 2016 at 17:26 | #130

    @Theguyuk
    Problem is all company’s do it, it’s only a public scandal through the media (as with volkswagen) that can have a influence (fines). Does this mean they don’t do it anymore? Probably not, they just search for another way to cheat.
    But I agree with you: Fraud is fraud even if it’s only 1 dollar.

  130. September 6th, 2016 at 17:29 | #131

    @Theguyuk
    Install CPU-Z, check the second or third tab, and you should see the “screen resolution”.

  131. tkaiser
    September 6th, 2016 at 18:23 | #132

    @Theguyuk
    Food standards? Do you consider clockspeeds an ‘ingredient’?! I think the main mistake people make is still looking at the wrong numbers. SoCs/CPUs running at lower clockspeeds are able to outperform those running at higher clockspeeds. Why? Since efficiency and since performance matters and not clockspeeds. Seems we must use the car analogy: You want mph (performance) but are only talking about rpm (clockspeed) and totally forget the gearbox in between (SoC architecture, efficiency, software).

    And while most benchmarks are fooling users no benchmark has been fooled. The benchmarks reported the performance (mph) and these numbers remain the same even if assumed rpm are lower (which is good news since lesser rpm/clockspeeds higher efficiency).

    And this is nothing new, please look at the table above where the whole story originated from. There pcDuino8 Uno relying on H8 SoC is listed as cheating. This SoC is sold under 5 or 6 different names in the meantime, I tested with A83T last year. Hardware and kernel allow the SoC running even at 2.1 GHz but default budget cooling settings prevent clocking higher than 1.8 GHz. Unlike Amlogic Allwinner enables the user to adjust these settings easily so I did exactly that. Unlocked the 2.1 GHz by adding a new dvfs operating point at 1.1V VDD_CPUX and adjusting the so called THS and cooler_table settings.

    Gain in performance: ~15% for single threaded workloads. That’s nice. Gain in overall performance? Zero. Why? Since overheating/throttling prevents using any of the higher clockspeeds. When running demanding CPU workloads (eg. cpuburn-a7) on the SoC even with a large heatsink the CPU cores will throttle down to 1.4 Ghz or even below. Now add a fan or liquid nitrogen and you run into the next problem: these devices are powered through Micro USB and will then simply power off (Micro USB is crap with such devices, limited to 1.8A max and leading to severe voltage drops). So you need a solder iron to replace the Micro USB with a barrel jack and a 5V/4A PSU maybe just to realize that now the PMIC (power management IC) will be the bottleneck since it limits maximum current to the SoC. It’s either/or. You get 8 cores or 2 GHz but not at the same time. And that’s true for any more recent SoC in this segment. The maximum CPU clockspeed is irrelevant anyway regardless whether the SoC vendor is ‘cheating’ or just disabling these frequencies in his dvfs / throttling / cooling settings.

    I hope that the attention Amlogic now gets will result in the great ODROID community, Hardkernel and Amlogic working together to provide new reasonable dvfs settings making use of 2 GHz with 1 core, 1.75 GHz with 2 and 1.6 GHz with 3. And then ARM arrives where Intel was already a decade ago: Turbo Boost (the only reasonable way to combine what the CPU cores are able to run at with what the environment allows)

    And the whole issue is not about cheating but understanding and wrong expectations. People want high miles per hour but talk about the engine’s rotations per minute forgetting both about the gearbox in between and the use case (in the field you’ll make a mistake choosing the Lamborghini Huracán over the Nitro when looking at mph/rpm instead of what really matters)

  132. Theguyuk
    September 6th, 2016 at 19:43 | #133

    @cnxsoft
    Cheers and yes above 720 the MP2 Mali400 suffers terrible frame drop on the Remix Mini, I tested both 50 and 60 Hz screens at 720 and 1080. I also did the milk run 4 x at each resolution ( very boring! ).

    A separate problem my Bluetooth rubber keyboard does not map the stunt button on Remix mini but does on the Mini M8S s905. On mini m8s s905 supplied remote control and my Bluetooth keyboard work, ( remember I am using the free Amazon underground app which is mapped to the Amazon remote and Amazon game pad ).

    Also noticed flickering and graphics tearing in the intro video to Factory Shift on Remix mini.

    Cheers for explaining that cnx-software ☺

  133. Theguyuk
    September 6th, 2016 at 21:04 | #134

    @tkaiser
    Hello tkasier ☺

    I used food example to highlight that with out standards you don’t know what you get.

    Same as 100 dollars are 100 not 98 and to old tin cans.

    You mention power but countries have different electrical standards a USA charger is different to a UK charger. You need agreed defined terms and standards. A millimetre is a set size and not changeable.

    As to benchmark software they have been fooled they reported the score as x amount with Soc at 2GHz but the AMLogic can not do 2GHz, that reported speed was false and undetected.

    I hear your car analogy but its flawed because that’s not a true model of what’s happening in the Soc. Steve Furber, he designed the physical first ARM RISC explains that it is memory bandwith between CPU and memory, how many clock cycles executing a instruction takes and how little the different Soc parts waste time waiting on the next job.

    Yes more speed needs more power causing more heat. But if it takes a instruction 5 clock cycles, and memory has the next instruction ready to give in 2 clock cycles, you get bottlenecks.

    It is not just speed but how quickly you can complete a instruction. ( you may have a 2GHz CPU but if add 1 + 1 takes 20 clock cycles, and different 1.5GHz CPU that can do add 1+1 in 10 clock cycles, where clock cycles in both CPU take the same time).

    The most instruction efficient CPU wins the race.

    Hence ARM ( simple instructions ) beat Intel ( complex instructions ) at lower power needs and die size.
    No matter the gearbox used. The hardware in and outs effect elsewhere, it is the efficient in sync orchestration of the whole that matters.

    For any less technical people reading this see below
    ( that is a very simplified example the real world CPU is more complex, perhaps a broadband model helps better. Some parts of a Soc only have slow dail up communication, other parts have copper wire asdl, while a few lucky parts have fibre broadband. Yet they all have to work together to make the Soc work, which means some parts have to wait for their next instruction. While others are so quick they pile finished instructions up, cached. But blockages build up and over heat the connections so you have to tell some parts to wait or rest, till others catch up. It a analogy to help less technical people understand. )

    But just to make it harder the CPU and GPU both get hot because of more electrical power causing more work instructions to be delivered to be done. However the CPU and GPU work in the one office, called Soc, and the office keeps getting smaller and smaller while over hot CPU and GPU get pushed closer and closer together. And the amounts of instructions for CPU and GPU to do keeps getting bigger and bigger.

  134. NIPSZX
    September 7th, 2016 at 04:01 | #135

    Why are there so many comments on this page? I have been following CNX for 2 years and I have never seen so many comments. What is going on? Is this why the tech sector in the stock market went up so much today?

  135. wpcprez
    September 7th, 2016 at 04:03 | #136

    what does the tech sector have to do with cnx comments? people care because they feel cheated so they are commenting. I’m fine either way and if it can go 2ghz im happy, if not the boxes weren’t that expensive so oh well.

  136. blu
    September 7th, 2016 at 13:41 | #137

    thesandbender :
    I’m guessing the market for a ~$200 board with proper SATA, upgrade-able RAM, etc really isn’t there. Closest you can get is a Tegra TX1 which is $600 and doesn’t even come with a SATA port (just M.2 though you can add a controller via the PCI-Ex4).

    ClearFog ARMADA 8040 ($300 for all the features you wish) should be out next month. Mine is preordered.

  137. tkaiser
    September 7th, 2016 at 15:01 | #138

    @blu
    8040 is also headless, I would assume @thesandbender is interested in an OTT box solution that does not suck regarding storage (and as already written a few meters above: ARMADA 1500 might then get interesting but I’ve no idea if devices with these SoCs already exist).

    BTW: after reading the comments here and seeing so many people being focused only on irrelevant numbers I also wonder how many readers think clockspeeds would be an absolute performance indicator? Thinking an S805 clocking at 1.5 GHz would be as fast as a S905 since… 1.5GHz!

  138. Theguyuk
    September 7th, 2016 at 15:49 | #139

    @tkaiser
    A very valid point and question, we are lucky because cnx-software looked into the subject before in 2015 !

    http://www.cnx-software.com/2015/04/09/relative-performance-of-arm-cortex-a-32-bit-and-64-bit-cores/

    Now where are those A17 and A15 quad TV boxes running Android 5, 6, 7. Mind you there is power consumption and heat too, although a TV box has no battery to bother about..

  139. September 7th, 2016 at 16:48 | #140

    blu :
    ClearFog ARMADA 8040 ($300 for all the features you wish) should be out next month. Mine is preordered.

    I think the Inforce 6410Plus board looks like a good contender at $143, although the RAM is not upgradeable.

  140. tkaiser
    September 7th, 2016 at 17:42 | #141

    @Theguyuk
    Do you really think replacing one BS number with another does help understanding anything or educating users? Dhrystone has been invented 30 years ago. It might have been useful back then but in the meantime it’s just a nice test for compiler switches but does not tell that much about processor capabilities. But since people fear using their brain and compare numbers without meaning instead questionable DMIPS scores are collected over time and used. Just in case you want to get the idea how stupid DMIPS/MHz ratings based on different compilers and optimizations are: http://www.brendangregg.com/blog/2014-05-02/compilers-love-messing-with-benchmarks.html

    In the meantime even I got that it’s useless to continue. Only people that do not care about performance at all are affected by the ‘2 GHz problem’ since they believe clockspeeds are an absolute performance indicator, don’t get that more recent CPUs benefit from improved CPU design and better instruction set (which makes them faster at the same clockspeed) and that while CPU cores are able to run at high clockspeeds the simple reality (temperatures / consumption) prevents that.

  141. paul
    September 7th, 2016 at 18:47 | #142

    @tkaiser
    Like your post. I just commiserate the difficulty of finding one or more useful CPU benchmarks that one can use to compare ‘universally’ over many years – perhaps impossible. For some years I try to rely on notebookcheck.net rankings for Intel/AMD.

    Even worse scenario appears to be a lack of standardized Range/Speed benchmarking of WiFi performance, both on miniPCs and other platforms. Perhaps you have some thoughts on that ?

  142. September 7th, 2016 at 18:53 | #143

    @tkaiser The title of that blog entry is misleading and though its primary conclusion is correct (compilers should match for benchmarks) the statement that compilers “mess with benchmarks” is just wrong (at least for gcc and clang). The compilers are doing their job, which is to look for inefficient code and optimize it. A better title would have been “compilers expose poor benchmarks”. i.e. if your benchmark just adds 1 to a variable one million times and then doesn’t use the variable, the compiler is going to strip that “dead” code out. They do the same thing with branches that are never called, etc. That’s one of the reasons you have different optimization levels (e.g. -O, -O2, -O3), sometimes the compiler is not as smart as it thinks it is and removes something that maybe referenced indirectly.
    There’s also a limit to how much optimization improved microcode can do. If an ARM register to register add instruction takes one cycle to execute there’s nothing that can be done to make that instruction faster as one clock cycle is the minimum granularity. The only way to get more speed in that case is to increase clock speed. So clock speed does help but it is absolutely possible to have a 1.5ghz CPU that (on average) performs better than a 2.0ghz CPU with the same instruction set, but there are activities that the 2.0ghz will always be faster at, no matter what.

  143. September 7th, 2016 at 19:19 | #144

    @blu @RobBrownNZ Thanks guys, both boards look very interesting.

  144. tkaiser
    September 7th, 2016 at 19:38 | #145

    @thesandbender
    Whether the title is correct or not I don’t care that much. It’s all about that passive benchmarks produce only numbers without meaning (you have to check the benchmark itself what it actually does and whether this translates to anything useful if you have a look at reality. And while it seems like using the same bench for over 30 years since ‘consistent results’ quite the opposite is true).

    And of course if CPU cores can clock higher performance should improve. In an ideal world or environment. A TV box or an SBC is the opposite by design (when we started to use sane settings with Allwinner’s H3 for example and reduced clockspeeds and voltages overall CPU performance increased!). The problem is that people only look at bullshit numbers like clockspeed. An ODROID C1/C1+ (or S805 in general, Cortex-A5, whatever DMIPS/MHz ratio) needs 150,75 seconds for running the usual/useless sysbench cpu test with 20.000 settings. An ODROID-C2 (or any S905 device) needs just 5.7 seconds. Both SoCs running at 1536MHz clockspeed.

    These sysbench scores are also absolutely useless to compare different systems but at least all those people now whining they get less clockspeeds might get the idea that this is pretty much irrelevant when you take software optimization, CPU architecture and other factors like memory bandwidth/access into account. So the people that think S805 and S905 perform identical since ‘1536 MHz’ and think they need the additional 0.5GHz might be able to get at least a feeling that they’re looking at BS numbers since when sysbench numbers are trusted blindly S905 would be over 25 times faster than S805 running at the same clockspeed (again: the benchmark is the problem, calculating prime numbers is nothing a TV box has to do).

    The whole problem is just that people love numbers without meaning. Be it clockspeeds or silly benchmark numbers (I had a look yesterday how the usual Antutu ‘benchmark’ test happens: test execution way too short, different engines tested separately instead of together, so Antutu produces also only numbers without meaning since throttling is not considered at all it seems?)

    And temperatures/consumption is the problem and not clockspeeds. I’m pretty confident that S905 is able to run at 2.0 GHz. But only in a lab environment and not in a TV box or on an SBC (powering problems, thermal problems). I already tried to outline that before referencing Allwinner’s H8/A83T. This SoC is not able to run even mild workloads clocked with more than 1.2 GHz but people do not feel cheated since it’s possible to run single-threaded workloads at higher clockspeeds. Maybe Amlogic just need to do that now (invest a lot of work to implement ‘Turbo Boost’ on ARM) and the clockspeed fanatics are happy again?

  145. mo123
    September 7th, 2016 at 20:08 | #146

    Lots of Amlogic S905 boxes also don’t have a full 2GB RAM.
    But it is probably manufacturer and RAM supplier dependent.
    If you check you will see something in the region of 1750-1785MB RAM.
    That is the experience on a MiniMX S905.

  146. September 7th, 2016 at 20:15 | #147

    @mo123 Nothing nefarious there, the “missing” ram is allocated for the video card and firmware to use. With some boards you can allocated how much memory is allocated to the video card, e.g. if you’re using it headless you can set it to the minimum to free up more RAM. This applies to almost all SoC systems. You’ll see the same thing on any x86 systems using an IGP, some of the system RAM is set aside for the IGP (generally you can control how much in the BIOS).

  147. JotaMG
    September 7th, 2016 at 21:25 | #148

    Hey guys…
    the giveaway has ended so you don’t have to enter any more comments…
    just joking! hehe

  148. mo123
    September 7th, 2016 at 22:06 | #149

    @thesandbender
    I thought it might be used for the VPU/GPU.
    But it is just strange since Rockchip devices report the full 1.99GB RAM available to be used by apps.
    Not sure how it works on their devices.

  149. Theguyuk
    September 7th, 2016 at 22:08 | #150

    @thesandbender
    You will never get away from compiler difference on Reduced Instruction Set Chip by design RISC put the work back on the compiler, code writer. If I recall right.

    And Intel, AMD chips get to complex, eat energy for fun and burn. Putting the complex instruction on the chip use more clock cycles per instruction.

    The quick fix for over heating would be add a system board bus to the Soc and split the GPU, VPU back to a separate chip with own memory. But Bill Of Materials and power use climbs.

    So cheaper to under clock a 2GHz Soc to 1.5GHZ and passive cool. ( I am not saying that is what AMLogic have done )

  150. blu
    September 7th, 2016 at 22:26 | #151

    Theguyuk :
    @thesandbender
    You will never get away from compiler difference on Reduced Instruction Set Chip by design RISC put the work back on the compiler, code writer. If I recall right.

    Absolutely nothing to do with RISC/CISC.

  151. September 7th, 2016 at 22:34 | #152

    @Theguyuk The line between CISC and RISC has blurred a lot. The initial idea of RISC was to have a minimal amount of instructions and push complex instructions, instruction ordering etc. onto the compiler (as you described). The original MIPS chip had something like 70 instructions (the Berkeley RISC chip had 32!). The latest iteration of AArch64 has ~200 I think (including NEON and Crypto). Also, the Intel “Core” line of chips are actually “RISC” chips internally. The chips only support a smaller set of operations referred to as microcode by Intel. Part of the pipeline is decoding the CISC x86 instruction set into the microcode (and that’s one of the reasons SMT works better on Core chips than “HyperThreading” did on NetBurst chips… the processor can schedule a arithmetic microcode from one “CISC” instruction while it’s it’s stalled out on a load microcode from another “CISC” instruction. Core was never actually intended to be Intel’s primary architecture, Intel’s Israeli team was tasked with making a mobile cpu that wasn’t as power hungry and the CISC/RISC hybrid is the solution they came up with. Meanwhile, the primary team found out they couldn’t push the NetBurst architecture any faster (clockwise) and Intel switch everything over to core architecture (and the rest is history). The Microcode solution also allows Intel to “fix” broken instructions in some cases (not all)… that’s whats in the “Intel uCode” package on linux distros.

  152. DarthMaul
    September 8th, 2016 at 22:18 | #153

    If you’ve only spent 30 or 40 dollars on an Android TV box, hen it’s no big deal you got an absolute bargain for your cash providing you purchased one of the better cheap china devices that now have good LE support like MiniMX and some nexbox models etc.

    When your spend 100 dollars or more on the expensive brands sometimes with less RAM and flash than the cheap china boxes , you should be angry that you have been short changed and not recived the spec you paid for. Minix, Tronsmart, Wetek etc… You pay more money for supposedly more trustworthy established brands with better support and testing….

  153. theguyuk
    September 9th, 2016 at 14:00 | #154

    @AMLogic

    There is hope you 1.5GHz limited Socs might run at 2GHz yet

    Extreme temperature specialist Vorago Technologies says its VA10800 microcontrollers have now been operating successfully in a temperature controlled oven for more than 4000 hours at 200˚C. During this time, the CPUs – based on the ARM Cortex-M0 core – have performed more than 7200trillion sequencial error free operations and exhibited consistently low current consumption.

    😈😈😈😈😇😀

  154. tkaiser
    September 9th, 2016 at 14:20 | #155

    @theguyuk
    That says it all. Comparing an absolutely limited microcontroller running at 50MHz/200°C with TV box SoCs (read as: more than just CPU cores) at 1.5GHz or above where environmental conditions prevent them running at their theoretical maximum. Ignoring completely the relationship between voltage, temperature and clockspeed. I think I already posted this link here in the comments: http://asic-soc.blogspot.de/2008/03/process-variations-and-static-timing.html (I know, it’s useless, people affected by GHz hype don’t want to get the basics. You really think every chip is the same as long as ARM is written on it?)

  155. September 9th, 2016 at 20:58 | #156

    Details about the new Amlogic S905 / ODROID-C2 frequencies options with the latest Linux 3.14.77-80 kernel @ http://odroid.com/dokuwiki/doku.php?id=en:c2_set_cpu_freq.
    Short summary:
    * 1.920 GHz stable with one core, and can boot with 2 cores.
    * 1.656 GHz stable under all conditions.

    Frequencies and cores count can be selected via boot.ini text file.

  156. blu
    September 9th, 2016 at 21:39 | #157

    cnxsoft :
    Details about the new Amlogic S905 / ODROID-C2 frequencies options with the latest Linux 3.14.77-80 kernel @ http://odroid.com/dokuwiki/doku.php?id=en:c2_set_cpu_freq.
    Short summary:
    * 1.920 GHz stable with one core, and can boot with 2 cores.
    * 1.656 GHz stable under all conditions.
    Frequencies and cores count can be selected via boot.ini text file.

    Not bad, actually (for a 28nm chip with hypothetical limit of 2GHz)

  157. Theguyuk
    September 11th, 2016 at 15:52 | #158

    For those not technical but wanting to try playing with, changing some Android setting not available in the provided setting, I have tried a few Google play apps that can change things. I only used free options or advert passed.

    Warnings first !!!!!!!!!!!

    I am not responsible for any damage caused, I give nor claim any warranty or technical ability on my actions or part. Using, trying anything I write about is clear indication you take your own responsibility for your actions.

    AMLogic s905 -s912 is locked by vender to 1.5GHz I am not aware you can change that on a Android TV box!.

    I only tried following in super user mode with mbox M8 s802 2gb, mbox mini s905 2gb.

    1, Antutu CPU Free, let’s you change governor and lower CPU speed.

    2, Quad-Core Processor Booster, by Pace. Use the basic free option and apply. They also do versions for other multiples of CPU.

    3, With Kernel Adiutor you can tweak and monitor things like:

    CPU (Frequency, Governor)
    I/O Scheduler
    Kernel Samepage Merging
    Low Memory Killer (Minfree settings)
    Virtual Memory
    Flash/Backup
    Build prop Editor
    Recovery (Flash, Wipe)
    Init.d editor
    Saving Profiles

    Any effects are subjective, I am not responsible for any Damage caused. Using any of these apps is your own choice…

    No doubt I will get flamed and moaned at by some but to other enquiring minds enjoy.

    PS CPU-z app claims to show CPU temp, I have no means or knowledge if CPU-z reading is accurate.

  158. tkaiser
    September 19th, 2016 at 18:50 | #159

    Too funny. While the whole SBC world complains about ‘2 GHz cheating’ by Amlogic/Hardkernel the most careless SBC vendor ever uses still the 2 GHz when producing his usual fake comparison charts:

    http://forum.banana-pi.org/t/banana-pi-bpi-m64-vs-raspberry-pi-3-vs-odroid-vs-pine64/2070

    They also don’t get it that their own product has VPU support (at least when running Armbian on it — any Pine64+ image will do it), that Android 6.x and 7.x is available for A64 in the meantime, that Pine64 has no separate Micro USB OTG port and so on. Could be funny if not the very same careless person producing this marketing chit chat would also be responsible for their ‘technical documentation’. At least it’s no wonder that their docs/specs never are correct 🙂

  159. Abraham
    December 19th, 2016 at 19:14 | #160

    Hello
    Gpu Amlogic S905X (Mali-450 GPU up to 750MHz+)
    4.2.2 Images show?

    There decoding of images 4.2.2?

  160. klimek stanislaw
    March 19th, 2017 at 01:49 | #161

    Hi, I have a question I bought something I zoomtak UPLUS No Yes enter SETUP in Kodi and broke my complete beginners

  1. No trackbacks yet.