Home > Allwinner H-Series, AMLogic, Rockchip RK33xx > Amlogic S905X vs Rockchip RK3328 vs Allwinner H6 Processors – Benchmarks & Features Comparison

Amlogic S905X vs Rockchip RK3328 vs Allwinner H6 Processors – Benchmarks & Features Comparison

November 27th, 2017 Leave a comment Go to comments
Orange Pi Development Boards

Rockchip, Amlogic and Allwinner are all battling for the lower and mid range segment of the TV box market, so it may be interesting to compare their solutions. We won’t look into the ultra low-end market with 32-bit ARM Cortex A7 processor, but instead compare some of the recent quad core 64-bit ARM Cortex A53 processor for 4K HDR TV box from the company with respectively Amlogic S905X, Rockchip RK3328, and Allwinner H6 SoCs.

We’ll compare some of the benchmarks obtained with Android TV boxes, as well as other features like video support, USB and Ethernet interfaces.


Let’s start with results for popular Android benchmarks: Antutu 6.x, Vellamo 3.x, and 3DMark Ice Storm Extreme v1.2 with results obtained from 3 TV boxes: Mini M8S II (Amlogic S905X), A95X R2 (Rockchip RK3328), and Zidoo H6 Pro (Allwinner H6). A score is highlighted in green is there’s a clear winner, and in red for a clear loser.

Amlogic S905X Rockchip RK3328 Allwinner H6
CPU (1) Quad core Cortex A53
@ 1.51 GHz
Quad core Cortex A53
@ 1.51 GHz
Quad core Cortex A53
@ 1.8 GHz
GPU (2) ARM Mali-450MP3 ARM Mali-450MP2 ARM Mali-720MP2
Antutu 6.x
Overall 33,553 33,117 40,467 / 36,957 (2)
3D (1920×1080) 3,099 1,475 6,292 / 2,782 (2)
UX 12,365 16,426 13,360
CPU 12,438 10,486 16,395
RAM 5,651 4,730 4,420
Vellamo 3.x
Metal 910 937 930
Multicore 1,491 1,464 836 (3)
Browser 1,855 (Browser) 1,943 (Chrome) 2,546 (Browser)
3DMark – Ice Storm Extreme v1.2
Total score 4,183 2,252 3,951
Graphics score 3,709 1,871 3,643
Physics score 7,561 7,814 5,608

(1) Those are the frequencies reported by CPU-Z, and the actual maximum frequency may be different. For example, it appears Allwinner H6 can only run at 1488 MHz in a sustained manner, and possibly only reach 1.8 GHz during short bursts (TBC).
(2) Allwinner H6 is the only SoC to include a GPU supporting OpenGL ES 3.1, which means it is the only one to complete Marooned 3D graphics test (Antutu 3D test has two 3D benchmarks), and the other boxes just got zero since it did not run. So I’ve included two scores for overall and 3D Antutu results: actual score / score minus Marooned benchmark.
(3) Vellamo multicore had a warning on Zidoo H6 Pro, so it may not represent the actual performance of the device.

Allwinner H6 has a slight advantage, but during use it will be really hard to notice the difference between TV boxes with one of those processors, and other factor like RAM capacity and storage performance will have more influence.One exception is 3D performance, as Rockchip RK3388 is clearly slower here, and I could notice it while playing games.


But SoC performance is only one side of the equation, so let’s have a look at some of the features from the SoCs, which may or not be implemented in some TV boxes. For reference I also included USB 2.0 or 3.0 storage (HDD NTFS partition), and Ethernet performance numbers. Those numbers may vary a lot with further software optimization, configuration tweaks, so they should only be used for reference. I used the same TV boxes as for the benchmark section, except for Gigabit Ethernet relying instead on iperf results from ROCK64 development board (RK3328) and K1 Plus (Note S905, no X, for reference only, but in my experience all Fast Ethernet interfaces have about the same performance), and NEXBOX A95X for the USB storage performance.

Amlogic S905X Rockchip RK3328 Allwinner H6
– 4K 10-bit HEVC Up to 60 fps
– 4K VP9 Up to 60 fps
– 4K H.264 Up to 30 fps (8-bit only) Up to 30 fps (8-bit and 10-bit) Up to 30 fps (8-bit only)
USB 2.0 / 3.0 USB 2.0 USB 3.0 USB 3.0
– A1SD Bench (R/W) 37/37 MB/s 94.52/90.73 MB/s 59.07/42.12MB/s
Ethernet 10/100M only Gigabit Ethernet MAC Gigabit Ethernet MAC
– iperf (full duplex) 91.6/91.8 Mbits/s 815/344 Mbits/s 758/350 Mbits/s
RAM Capacity (Max) 2GB 4GB 2GB
Misc  TS, Smartcard interface TS, Smartcard interface, PCIe

I did not include audio, as all those SoC are supposed to support Dolby TrueHD and DTS HD audio codec pass-through. but implementation varies a lot between devices.

Amlogic S905X is the weakest of the lot based on the two tables above, but it’s also the cheapest SoC among the three, and in my experience, one with the best support in Kodi, for example. Rockchip RK3328 is not much more expensive, and have many benefits, except when it comes to 3D graphics performance, but it usually only matter if you plan to play games on the platform, the GPU is usually good enough for user interfaces. Allwinner H6 has more interfaces, a Mali GPU with OpenGL ES 3.1 and OpenCL support, and lightly more interfaces. The few devices that are based on the Allwinner processor are currently quite more expensive with all other features being equal.


  1. user 56
    November 27th, 2017 at 16:37 | #1

    Hope we get to see the 905W in the test group as well

  2. November 27th, 2017 at 16:41 | #2

    @user 56
    S905W should be about the same as S905X, except video support is limited to 4K @ 30 fps.

  3. Jakub Šimo
    November 27th, 2017 at 16:59 | #3

    Does it really matter how they are powerful if they have partial/non-existant support for things like OpenELEC/LibreELEC?

  4. Golgot13
    November 27th, 2017 at 17:00 | #4

    Surprise to not see the Hi3798 chipset
    To my mind Hi3798CV200 is the KILLER for chipset developped for video decoding market
    If you want the best of the best: Hi3559A, 8K video decoding in H265 AND the NNA => it could be used on SuperComputer

  5. blu
    November 27th, 2017 at 17:08 | #5

    Good comparison! Just keep in mind that even though not all GPU benchmarks may care about it, Mali-720 is the much more modern GPU — GL ES 3.x is only one of the benefits — OpenCL is another; basically, if you care about graphics (or GPGPU) Midgard’s advantages over Utgard are substantial. That said, of course you need all drivers/userland stacks in place, so you need to be careful about those as well. /linux warning

    ed: oh, you said all that at the end : )

  6. November 27th, 2017 at 17:14 | #6

    Other SoCs like Hisilicon Hi3798CV200 and Realtek RTD1295 are pretty great, but most device are in a different market segment, over $100, or in some cases over $200.

    I understand Hi3559A is mostly a camera SoC. You won’t find it in TV boxes, as AFAIK, some features like Dolby or TrueHD are not supported at all. If the info I got is true the chip by itself costs $50, so again, a different price category.

  7. theguyuk
    November 27th, 2017 at 18:05 | #7

    All hangs on what you want to do with the device you are buying.

    RK3328 could have the upper hand if software starts using the full 4GB ( needs a must have or killer application ). Most user reviews from TV watching user like some of the cheap boxes based on SoC.

    S905X Have more mature software use ( ? ). Sellers have more S905X experience on flaws. However Amlogic fib about true Hz of CPU

    Allwinner always screw the software and prior, it has been the community that sort it out, where they can get true hardware details.

    Am I right in thinking it is Armbian, RetroOrange Pi, Lakka and lately Friendlyelec who’s software has made prior H2, H3 Allwinner SoC actually useful? Orange Pi’s own software support, still stinks from what I read in users comments.

    Just IMHO

  8. tkaiser
    November 27th, 2017 at 18:19 | #8

    theguyuk :
    Am I right in thinking it is Armbian, RetroOrange Pi, Lakka and lately Friendlyelec who’s software has made prior H2, H3 Allwinner SoC actually useful?

    Wrong, look at http://linux-sunxi.org/ instead.

    @cnxsoft: Based on past experiences with Allwinner I would be really surprised if H6 storage and network performance are that low. Similar as with RK3328 where we get with more up to date Linux images (that take care of the necessary settings unlike vendor Android) Gigabit Ethernet fully saturated I would expect the same with H6 later (talking about +900 Mbits/sec in both directions). And if Allwinner licensed their USB3 implementation from Synopsys we’ll hit there the magical 400 MB/s USB3 SuperSpeed barrier too. And then there’s the single PCIe 2.x lane which makes H6 really interesting in case we get inexpensive dev boards with this SoC.

  9. theguyuk
    November 27th, 2017 at 19:03 | #9


    Followed your link, found this informative, thank you 🙂

    ” The sunxi community is a very active and advanced one. For more information, check our linux-sunxi community page. It explains how to get help and how to contact us, and how you can contribute as well.

    Allwinner does not actively participate in or support this community. In fact, it is violating the GPLv2 license in several ways and has so far not shown willingness to resolve this. “

  10. TC
    November 27th, 2017 at 22:14 | #10

    Rockchip for Anime!

  11. November 27th, 2017 at 22:28 | #11

    It would be great also to compare Linux support for the listed SOCs not just the hardware

  12. jernej
    November 27th, 2017 at 22:39 | #12

    H6 BSP kernel has much less licensing issues (some were reported and will be probably fixed in next release). Atop of that, it seems they want to help community according to this: https://irclog.whitequark.org/linux-sunxi/2017-11-14#20543979

  13. Tiago Marques
    November 28th, 2017 at 01:26 | #13

    Not really surprised that the RK3328 does so badly in CPU tests, the thing is implemented with an even paltrier 256KB of L2 cache (vs 512 on most A53 quads). Fine for NAS, I suppose, but something to avoid otherwise.

  14. Golgot13
    November 28th, 2017 at 07:03 | #14

    Yes Hi3559/Hi3559A was not developped for media player market, but for AR/VR on SoC camera
    But like media player, there are Video Output Interface, with one HDMI 2.0 output
    About audio format support, Dolby decoding is based from ARM/NEON optimization (Google: Dolby + Ittiam + ARM)

    About linux, Hisilicon give full linux support, so with hardware decoding available (and encoding 😉 ) on SDK

    Hi3798 is the first chipset with all 10bit certification support ( Dolby Vision, HLG and HDR10 )
    I listen AV1 codec decoding is in development on Hi3798CV200 …

  15. November 28th, 2017 at 09:13 | #15

    I performed same tests on VideoStrong KII Pro (S905) 2GB/16GB
    with custom firmware (Android 7.1.2 built from latest Amlogic SDK; ATV version; 320 DPI)

    Here are results from single run of each benchmark:

    Ice Storm Extreme
    Total score: 4,019
    Graphics score: 3,417
    Physics score: 10,492

    Antutu Benchmark:
    Overall: 37,352
    3D (1920×1080): 2,923
    UX: 14,949
    CPU: 14,315
    RAM: 5,165

    Metal: 1,085
    Multicore: 1,867
    Browser: 2,419

    • Markocc
      November 28th, 2017 at 17:42 | #16

      Hi Stane can you give me link to that firmwere,ty-hvala unapred 😉

  16. November 28th, 2017 at 09:24 | #17

    I found that S905 is usually a bit better than S905X too: https://www.cnx-software.com/2016/07/31/amlogic-s905-vs-s905x-benchmarks-comparison/

    I still used S905X for comparison as all three support 4K HDR and VP, and it seems many recent devices use S905X (or derivatives) instead of S905.

  17. November 28th, 2017 at 10:04 | #18

    Software support is always in flux, it’s very hard to keep track of it.

  18. November 28th, 2017 at 10:05 | #19

    What does “Allwinner BU3” mean in the IRC chat? BU3 = Business Unit 3, or something else?

  19. jernej
    November 28th, 2017 at 13:14 | #20

    Yes, BU means Bussiness Unit. I forgot which one is responsible for which SoCs. I guess it is safe to say BU3 is responsible for H series (media box line).

  20. November 28th, 2017 at 16:06 | #21

    Thanks. Finally I got some more information. There used to be 5 business units, but now only 4:
    * BU1 – R-Series
    * BU2 – A-Series and VR-Series
    * BU3 – H-Series and V-Series
    * BU4 – Automotive SoC

  21. theguyuk
    November 28th, 2017 at 16:57 | #22

    Cheers 🙂

  22. mdel
    November 28th, 2017 at 17:18 | #23

    your rk3328 Gigabit iperf speeds are not what i get on my cheapo (33e) Alfawise Z28 Pro (rk3328 2G/16G), i do get >850Mbps bidirectional.

    That z28 Pro box Gbe implementation is pretty much identical to the rock64 one (rk3328 GBE MAC with same external PHY).
    The one thing completely fucked up on that box is that they didn’t wire SDMMC0 for the sd card slot so you can’t boot from sdcard.. but then you have rockchip maskrom mode (pads on the board) which lets you flash the emmc directly through usb..

    Maybe your speeds have something to do with rock64 Gbe problems, i don’t remember if this was a real thing and if it’s been fixed since then.

    i’ll also say that there’s a massive 3mm(?) gap (drilled pcb) around the ethernet port on my board, it is specified in rockchip datasheets, i’m not sure what kind of high voltages they were expecting there but i’ve never seen anything like that..

    – One comment on a feature not described in those tests : Cryptoengine implementation and support, or lack of..
    if you’re planning on using those arm chips for encryption (openvpn, luks disk encryption..) you should be very careful about which chip you choose, and basically what i would say is avoid s905 (i don’t know about s905x but i suspect i’ll be the same).
    rk3328 has Cryptoengine support in android 7 (haven’t checked linux yet), s905 does not in the androids (5 or 6?) i’ve seen on my tv boxes, nor in the linux kernels available.
    S905 can only achieve around 100Mbps in openvpn (single cpu thread), rk3328 has no problem saturating the gbe link with its CE extension.
    Some android benchmarks do test AES speeds.

  23. tkaiser
    November 28th, 2017 at 17:38 | #24

    S905X has ARMv8 AES crypto extensions just as RK3328 (and since H5 also features them I would assume H6 does the same — no idea why in reviews always /proc/cpuinfo is missing). Currently the only known two ARMv8 SoCs missing those crypto extensions are S905 and Raspberry’s BCM2387.

    Wrt Gigabit Ethernet and USB3 the numbers here are made with vendor Android builds and therefore simply suck. With appropriate settings on RK3328 we get +900 Mbits/sec network speed and almost 400 MB/s with USB3 SuperSpeed and I would assume it’s the same with H6 and Linux using sane settings and a more recent kernel later (current H6 BSP relies on a horribly outdated 3.10.65 kernel)

    The above storage and network performance numbers are only interesting for TV box consumers who have to rely on crappy vendor Android builds (relying on outdated kernels like 3.10 and 3.14) and have no meaning for anything else (especially using these SoCs with Linux).

  24. tkaiser
    November 28th, 2017 at 18:42 | #25

    Thanks to Icenowy we now know that H6 also supports the ARMv8 crypto extensions according to /proc/cpuinfo:

  25. blu
    November 28th, 2017 at 21:23 | #26

    Let’s hope no vendor releases another A53 without those ever again.

  26. tkaiser
    November 28th, 2017 at 21:53 | #27

    I don’t really care as long as there are a few SoC options available with good Linux support where I can tick the following checkboxes:

    – Gigabit Ethernet (or even 2.5GbE)
    – very fast storage (USB3, good SATA* or PCIe)
    – ARMv8 crypto extensions
    – support for at least 2 GB DRAM

    BTW: Pine people will test next week whether H6 combined with an 4GB LPDDR module works or not (though only 3GB will be adressible).

    * ‘Good SATA’ is meant as neither Allwinner SATA (A10, A20, R40, V40) nor i.MX6 SATA (both implementations are way too slow)

  27. GanjaBear
    November 28th, 2017 at 23:22 | #28

    cnxsoft :
    @user 56
    S905W should be about the same as S905X, except video support is limited to 4K @ 30 fps.

    Except, S905W is actually capped at 1.2 GHz.

  28. November 28th, 2017 at 23:44 | #29


    when/if firmware will be ready for release I will do it. At the moment S905 does have video playback issues on Nougat firmwares. I was discussing issue with S905 and Nougat with Amlogic engineer and we don’t have solution yet (it’s SDK issue).

    • Markocc
      November 29th, 2017 at 03:27 | #30

      Ok,if you have stable andriod 6 for kii pro i will be happy,amlogic nougat is beta for all devices…tnx

  29. November 29th, 2017 at 09:15 | #31

    I forgot about that, but I’ve found your earlier comment: https://www.cnx-software.com/2017/08/14/tanix-tx3-mini-tv-box-is-powered-by-amlogic-s905w-soc/#comment-545994

    Amlogic is a funny company.
    Amlogic S905/S912/S905X initially advertised at 2.0 GHz max, oops only 1.51 GHz
    Amlogic S905W initially advertised at 1.5 GHz max, oops only 1.2 GHz.

    • Markocc
      November 29th, 2017 at 14:06 | #32

      I think amlogic s905w can go up to 1.5ghz easy(they disable that)only mecool amlogic s905w firmwere have ui 1080p..

  30. tkaiser
    November 29th, 2017 at 14:25 | #33

    cnxsoft :
    Amlogic S905W initially advertised at 1.5 GHz max, oops only 1.2 GHz.

    But it’s the same with S905X and S912 too. With unmodified firmware they’re running with multi-threaded stuff at 1.2 GHz max while reporting wrong/bogus cpufreq values back to the kernel all the time (see comments in Khadas VIM2 thread). Only Hardkernel possesses a firmware for S905 on their ODROID-C2 that reports real CPU clockspeeds and not faked ones as everywhere else.

  31. tkaiser
    November 29th, 2017 at 16:24 | #34


    Just as a quick reminder: Amlogic’s bl30.bin bootloader firmware controls cpufreq/dvfs behaviour and seems to implement something comparable to Intel’s Turboboost (the less CPU cores are active the higher the maximum clockspeed can be which is not the worst choice to improve single thread performance within a given thermal and consumption budget). But the problem with Amlogic’s firmwares is that unlike almost everywhere else they do not report the real clockspeeds the CPU cores are running at but only fake values.

    That’s situation with S912 and at least with S905X it’s reportedly the same (max clockspeeds lower all the time but especially when all 4 cores are active):

  32. L.Peter
    November 29th, 2017 at 16:37 | #35

    I tought only chinese soc manufacturers fake frequencies.
    I hope they will be punished for cheating like volkswagen…..

  33. tkaiser
    November 29th, 2017 at 17:11 | #36

    L.Peter :
    I tought only chinese soc manufacturers fake frequencies.

    Hmm… I’ve not seen this so far but my experiences are very limited (I played with this low-level stuff only on Allwinner, Amlogic, Broadcom and Rockchip SoCs so far).

    Allwinner and Rockchip ship with bootloaders and kernels where users (or device vendors) have full control. The kernels define maximum cpufreq and possible cpufreq operation points (OPP), the whole DVFS setup is adjustable (device tree settings) and reported clockspeeds are real. There’s always a mismatch between maximum cpufreq A (highest cpufreq OPP allowed by the kernel) and maximum cpufreq B (highest cpufreq allowed by DVFS — dynamic voltage frequency scaling — and thermal budget) but at least the CPU clockspeeds that the kernel reports at a certain time are real.

    So with an Allwinner or Rockchip box you get something that ships with a kernel able to clock up to 1.8GHz for example but almost always on a device where the vendor only allowed something lower: for example 1.4 GHz due to consumption and thermal constraints. That tools like CPU-Z, Geekbench and other benchmarks only show the higher value is not really the SoC makers problem but a stupid limitation that needs to be fixed in the respective softwares (it’s f*cking easy to walk through the available cpufreq OPP within milliseconds and report the highest real clockspeed instead of reporting BS as all Android tools in this area do)

    With Amlogic and Broadcom (BCM2386/2387) it’s different: here the whole DVFS/cpufreq stuff is part of a proprietary firmware that will be loaded at boot and controls the SoC’s behaviour. The kernel doesn’t even know at which clockspeed the CPU cores are running. But at least with Broadcom (RPi 2 and 3) there’s a tool to query the firmware directly and only this command tells the truth. If you happen to run an underpowered Raspberry Pi 3 (as most people do thanks to crappy Micro-USB powering) then while something demanding is running the kernel happily reports the CPU cores being at 1200 MHz while querying the firmware using

    will show the real clockspeed: 600 MHz (on RPi 2 it’s 900 MHz vs. 600 MHz of course). The firmware does so called ‘frequency capping’ but doesn’t tell the kernel (and almost every RPi monitoring tool out there shows also these BS numbers instead of the real ones). Same with throttling: if the SoC overheats clockspeeds are lowered but the kernel doesn’t know about and still reports running at the maximum.

    Maybe such a tool to communicate with Amlogic’s properietary firmware is available but honestly I don’t really care since it doesn’t matter that much anyway. Since real performance is something different than CPU clockspeeds. It’s just somewhat annoying to not be able to read out and control these clockspeeds directly in case you want to optimize settings (which is fun especially with Allwinner SoCs since they ship with horrible ‘factory settings’… but on platforms where a proprietary bootloader controls this it’s simply not possible since part of ‘the firmware’)

  34. GanjaBear
    November 29th, 2017 at 17:36 | #37

    cnxsoft :
    I forgot about that, but I’ve found your earlier comment: https://www.cnx-software.com/2017/08/14/tanix-tx3-mini-tv-box-is-powered-by-amlogic-s905w-soc/#comment-545994
    Amlogic is a funny company.
    Amlogic S905/S912/S905X initially advertised at 2.0 GHz max, oops only 1.51 GHz
    Amlogic S905W initially advertised at 1.5 GHz max, oops only 1.2 GHz.

    A correction to the linked comment – VP9 is actually supported in TX3’s firmware, but for some weird reason, all mention of it is absent from their promo materials.

  35. mdel
    November 29th, 2017 at 18:35 | #38

    tkaiser :
    S905X has ARMv8 AES crypto extensions just as RK3328 (and since H5 also features them I would assume H6 does the same — no idea why in reviews always /proc/cpuinfo is missing). Currently the only known two ARMv8 SoCs missing those crypto extensions are S905 and Raspberry’s BCM2387.

    So all Almogic S905 datasheets i could find (ie hardkernel S905_Public_Datasheet_V1.1.4.pdf) are false ?
    They do show a Crytpo Engine on their block diagram..

    i’ll get a s905x cheapo then (what about s905w, this is the cheap one right?), to see how it performs with openvpn, are there any box with decent Gbe ?

    Also can you quickly describe how you test usb3 SS bandwidth ?
    I can’t get stable SS high speeds (>250MB/s) on my rk3328 box, i’m using various usb3 sata adapters with decent SSD hard drives (evo 850 pro, HyperX Savage), tests only done in android at the moment.

  36. tkaiser
    November 29th, 2017 at 18:45 | #39

    Except, S905W is actually capped at 1.2 GHz.

    It seems to be even worse. Reported max cpufreq is 1.2 GHz but according to benchmarks numbers Amlogic’s bl30.bin firmware clocks down to 1000 MHz or even 900 MHz when all 4 cores are active (hard to tell from the Geekbench numbers since a lot of reported ‘CPU performance’ depends on memory bandwidth/latency too).

    This is is S905X based Le Potato: https://browser.geekbench.com/v4/cpu/3197390

    And this is S905W based Vorke Z5: http://browser.geekbench.com/v4/cpu/4934864

    We know from S905X that it’s running at around 1.45GHz with one CPU core active and around 1.2GHz when all 4 cores are busy (while cheating on us running at 1.5 GHz all the time of course).

    By comparing numbers with S905W it’s 1.2GHz with one busy task and below 1.0GHz on all CPU cores with same loads. AES scores are not affected (which means a) that S905W not so surprisingly also features ARMv8 crypto extensions and b) the AES engine does not suffer in the same way from frequency capping as stuff running on the CPU cores does since ‘4 threads’ numbers are pretty exactly 4 times higher than single thread numbers).

    Looks like S905W is an S905X that didn’t fully pass the fab tests (memory controller instable at highest bandwidth so not suitable for highest resolution video and/or CPU cores too instable at highest clockspeeds) so instead of throwing the S905X die away it will become a S905W with limited memory bandwidth and lower CPU clockspeeds.

  37. November 29th, 2017 at 19:04 | #40

    It has been discussed @ https://www.cnx-software.com/2017/10/26/linux-benchmarks-intel-j3455-apollo-lake-vs-z3735f-bay-trail-vs-rk3399-and-other-arm-platforms/#comment-548150

    S905 has a crypto engine, but it’s proprietary, it does not support ARMv8 crypto extensions.

  38. tkaiser
    November 29th, 2017 at 19:29 | #41

    mdel :
    i’ll get a s905x cheapo then (what about s905w, this is the cheap one right?), to see how it performs with openvpn, are there any box with decent Gbe ?

    There are none with GbE and there never will be since S905X/S905W (most probably chips from the same wafer) do not feature GbE. Seems Amlogic is simply focusing here on what they do best: TV boxes with great codec support (and none of these devices needs GbE, same with USB3)

    To get close to 400 MB/s with RK3328 (and most probably also H6) you need 4 things:
    – storage that is able to exceed 500 MB/s sequential performance (most SSDs do not especially with low capacity or only for small amounts of data — see Samsung’s TurboWrite implementation since you’re using an EVO)
    – high quality UAS capable USB-to-SATA bridge and UAS support in kernel (no way with RK’s Android for example)
    – IRQ affinity settings that move away the USB3 interrupt from cpu0
    – large blocksize for tests (16MB or above)

    In other words: such a test is interesting to check the USB3 implementation’s limitations but is not that important for real-world scenarios where most of the times the storage will be the bottleneck (and in NAS scenarios limited network bandwidth of course)

  39. tkaiser
    November 29th, 2017 at 20:21 | #42

    Forgot before: there are two other things that are important for storage performance and this time it’s not about synthetical benchmarks but why often real-world Android storage performance sucks so much (see also numbers above, especially for the Allwinner H6 box).

    – cpufreq scaling settings: often a so called governor is used in Android that does not take care of IO activity and can lead to the CPU cores remaining at their lowest allowed clockspeed all the time slowing storage access down massively
    – filesystem in question: ‘Native’ filesystems like ext4 have a clear advantage since all the processing is way more efficient compared to ‘foreign’ filesystems like NTFS or ExFAT that usually are accessed through a FUSE layer (filesystem in userspace) which means a lot higher CPU utilization and lower performance at the same time

    While FUSE accessed filesystems usually come with a lot of CPU activity and therefore ramp up the CPU clockspeeds in any case with settings where each and every interrupt in the system is processed on cpu0 this first CPU core becomes a great bottleneck especially when for whatever reasons the kernel’s process scheduler doesn’t take care of this and sends also applications to this CPU core (been there, seen that. But with a smelly 3.4.x Android kernel years ago).

    And since I’m already babbling one final remark wrt ‘NAS setup’ especially when it’s about SMB. If Samba is used to provide Windows fileservices on all those weak ARM thingies it needs 5 simple Samba tunables that massively influence performance (up to 5 times faster with those few settings applied — if anyone is interested details with links in my ‘ODROID HC1 / HC2’ review in Armbian forum).

    And while all of the above performance talk isn’t really important for TV boxes since usually ‘fast enough’ for the use case I really wonder why vendor Androids do not take care a bit more about this stuff (since these performance gains due to better settings come at no cost) while on the other hand we should always have this in mind when ‘benchmarking different hardware’ as above: often it’s just different settings instead 🙂

  40. willmore
    November 29th, 2017 at 22:52 | #43

    How did you generate this table? I’d like to produce a similar table for my S905 based ODROID-C2. I’ve not seen any signs of lower clocks on differing # of cores before, but I didn’t have any test that I used to specifically used to look for that.

    It could be that the S905 is different from the S905W and S905X in that way? And S912, as well.

    When HK added the ability to set the clocks and # of cores active for the c2, there was no way to set different clocks for differing # of active cores. There was a setting for clock speed and one for # of active cores. And it enforced that by simply disabling any cores above the # enabled. So, it was more of a gate than a “Freq X for Y cores” kind of table. That might have been nice. Since the clock speed on that chip is limited to some extent by temperature, fewer cores *could* clock higher than the 4 core speed I have set.

    I don’t pick clocks that aren’t stable with all cores and with the harshest load I can manage because I prefer a machine that is unconditionally stable. You never know what workload will trigger a high load.

  41. tkaiser
    November 29th, 2017 at 23:28 | #44

    willmore :
    How did you generate this table?

    Simple sysbench runs with 1, 4 or 8 threads walking through cpufreq OPP (see last pages of ‘Armbian for Amlogic S912’ thread over in forum — there is the script I used). Since the sysbench cpu benchmark does not depend on anything else and runs completely inside the CPU caches with same binary on same platform it can be used to determine real CPU clockspeeds.

    With S905 at least on ODROID-C2 it should be different since Hardkernel got the bootloader sources from Amlogic after last year’s ‘excitement’ about Amlogic ‘up to 2 GHz’ cheating. But most probably the average S905 TV box out there still ships with a firmware controlling cpufreq/dvfs in a similar way as above (cheating about ‘2 GHz’, allowing ~1.5GHz with single threaded loads and dropping down to 1.2 GHz when all CPU cores are busy)

  42. willmore
    November 30th, 2017 at 19:53 | #45

    Ahh, okay, that thread. We already have my data for 1 and 4 threads. I’ll rerun it and set it for two threads. Then we should have all the data needed to make a table like that for the C2. I’ll post over in that thread.

    For those playing along at home: https://forum.armbian.com/topic/2138-armbian-for-amlogic-s912/

  43. tkaiser
    November 30th, 2017 at 22:43 | #46

    willmore :
    For those playing along at home: forum.armbian.com/topic/2138-armbian-for-amlogic-s912/

    Well, the thread is a bit boring if one is only interested in Amlogic cheating. The interesting part starts on page 17 there. Anyway: your numbers are based on most recent bl301.bin blob from github.com/hardkernel/u-boot/tree/odroidc2-v2015.01/fip/gxb (I always confuse this blob with bl30.bin). But even the older blob that faked the 1752 and 2016 MHz cpufreq OPP showed sane behaviour wrt single threaded vs. multi-threaded loads: sysbench numbers scale linearly with count of CPU cores so Hardkernel from the very beginning had a firmware blob for S905 that showed at least sane cpufreq behaviour up to 1536 MHz. No idea whether this also applies to u-boot/firmware used on typical S905 TV boxes.

    Speaking of those I did a quick search yesterday in Geekbench browser for S905 devices just to realize that there are some existent that seem to use either S905X/S905W in reality or there’s a new S905 with crypto extensions enabled (simple check: AES crypto scores not 25 or below 30 but above 350). Searching for ‘MXQ PRO’ for example you’ll find some results listing (faked) 2.0 GHz and an AES score of 25 (so most probably the real S905) but also results with 1.5 GHz and 1.2 GHz and great AES scores (maybe then an indication for S905X and the crippled S905W?). Unfortunately all 3 SoC identify themselve as ‘ARM implementer 65 architecture 8 variant 0 part 3331 revision 4’.

  1. No trackbacks yet.