Rockchip RK3588 specifications revealed – 8K video, 6 TOPS NPU, PCIe 3.0, up to 32GB RAM

Rockchip RK3588 is one of the most anticipated processors for the year on this side of the Internet with the octa-core processor features four Cortex-A76 cores, four Cortex-A55 cores, an NPU, and 8K video decoding support.

[Update December 2021: check out our post with the RK3588 datasheet for the latest details about the processor]

The roadmap shows an expected launch date in Q3/Q4 2020, but sadly the release date will be pushed back in the future. Having said that, the Rockchip Developer Conference (RKDC) is now taking place, and the company has put up a poster that reveals a bit more about the processor.

That means we now have more detailed Rockchip RK3588 specifications:

  • CPU – 4x Cortex-A76 and 4x Cortex-A55 cores in dynamIQ configuration
  • GPU – Arm Mali “Odin” MP4 GPU
  • AI Accelerator – 6 TOPS NPU 3.0 (Neural Processing Unit)
  • VPU – 8Kp60 video decoding support, 8Kp30 encoding support
  • Memory I/F – LPDDR4x/LPDDR5 up to 32GB
  • Storage – eMMC 5.1, SDIO, SATA 3.0 (multiplexed with PCIe 2.0)
  • Video Output
    • Dual HDMI 2.1 / eDP up to 8Kp60 or 4Kp120
    • Dual DisplayPort up to 4Kp60
    • Dual MIPI DSI output
    • Up to four independent displays
  • Camera – 48M (2x 24M) ISP with HDR and 3D NR support; multi-camera input
  • Audio – Microphone array support
  • Networking – Dual Gigabit Ethernet
  • USB – 2x USB 3.1 Type-C, 2x USB 2.0
  • PCIe – 4-lane PCIe 3.0, and 3x PCIe 2.0 (multiplexed with SATA)
  • Manufacturing process – 8nm LP

The company will provide support for Android, Linux, and a “domestic OS”. We should note that the GPU has changed from “Natt” family to “Odin” family. Rockchip is unable to disclose the GPU name (if it already exists) as it’s a new GPU family that has yet to be announced. We were informed that Rockchip designed its own NPU IP for RK3588, and did not use a third-party design like VeriSilicon NPU IP.

RK3588 specifications are pretty impressive, and the processor will be found in Arm computers, smart displays, edge computing & AIoT solutions, Arm servers, high-performance tablets, network video recorders, virtual reality headsets, and applications requiring multiple cameras and displays. We should however note that RK3588 sadly lacks UFS support, and relies on the slower eMMC 5.1 flash interface for storage. The launch is now scheduled for Q2/Q3 2021.

Share this:

Support CNX Software! Donate via PayPal or cryptocurrencies, become a Patron on Patreon, or buy review samples

109 Replies to “Rockchip RK3588 specifications revealed – 8K video, 6 TOPS NPU, PCIe 3.0, up to 32GB RAM”

  1. Amazing news, hopefully won’t have to wait a whole year to get some RK3588 SBC’s from Radxa and others. 8Core CPU, Up to 32Gb of LPDDR4x/LPDDR5 RAM, 8nm process, 3x 4-lane PCIe 3.0, and PCIe 2, got me sold already. I don’t think first RK3588 SBC’s will be cheap. IMHO, expect to pay at least $120 for minimal configurations.

      1. I hope they get the UART baud rate correct this time. The orange PI4 using Rockchip RK3399 had/has a non standard UART baud rate speed of 1500000.
        It also had powering related problems since the USB Type-C is not PD-compliant.
        I also hope they use a proper SATA chip and don’t use Ethernet via USB hub.

        1. The specs are for the chip itself, so SATA and 2x GbE are built into the chip. It would make no sense to increase the cost of the board by adding extra chips for SATA or GbE. I’m not sure I’ve seen an RK3399 board with USB GbE chip, unless there are two GbE ports.

          I’m not sure why the UART baud rate is an issue. Incompatibility with some USB to TTL chips?

      1. Awesome. Since I’m so commited to RockPi4C already, I definitely will be buying 1 or 2 Rock Pi5 from Radxa at that price.

      2. Will the software support be any good? For the Rock Pi S the software support seems to have been release and forget. It doesn’t seem to be possible to even use a version of u-boot that’ll compile with recent toolchains.

      3. > ROCK Pi 5, RK3588, 99$ for you

        LOL, and just one year of inflation blew that up to $129. Let’s see at which time of the year the announced shipping date of ‘Q2 2022’ will be. Maybe really ‘summer 2022’ as expected?

    1. Ironically, that would be quite cheap as I now look at this as an Apple M1 competitor. It supports twice the memory of Mac Mini, so the remaining factor is going to be the CPU clock speed.

      1. The M1 definetly has bigger performance in CPU and GPU.
        I think that only a SoC with a Cortex X1 could compete against the M1

        1. Even Cortex-X1 won’t be able to compete against the M1 because the M1 is much wider (8 vs 5), deeper, and equipped with huge and incredibly fast caches.

        2. So far I am not impressed with my M1 Mac Mini. However, I am coming from an ODROID-N2+ experience as reference. Maybe my poor experience with M1 is due to the fact that I use it for development work instead of running benchmarks.

          For those interested in quantifying data, how long did it take to do the point release OS update? How long did it take to install XCode?

          FWIW, either the OS is very poorly tested, or I got a bad device. Lots of issues “out of the box”.

          1. That’s why I’m waiting a few months to pick up mine.Hopefully those initial issues will be taken care of in net 5-6 months. I do own several Odroid-N2+’s as well.

          2. Thanks Thomas. Good compilation. Quite shocking seeing M1 efficient core give more perf vs A73 on Odroid N2+.

            Maybe a hint for Mr. Cubie to stack the ram with the SoC? Kidding..

        3. Apples to Apples. The cheapest M1 will be an Apple Mac Mini for $699, I can buy 3 Rock Pi 5’s@$99 each for less than half of that money and assign them to handle multiple tasks to fulfill my curent requirements. I may ended buying both M1 and RockPi5 to be able to run proper benchmarks and demistify M1 overall price/performance superiority.

          1. Wikipedia states Cortex-A76 max freq is 3Ghz in phones and 3.3Ghz in “tablets/laptops”. My expectation is that this RK3588 will clock in at 3GHz. At this speed, I doubt there will be much real world difference from M1 (unless you only run benchmarks). If RK3588 ends up being a 2.5GHz part, then its much less interesting (to me).

          2. I predict 2.4 unlocked so that vendors let testers like us figure how to stabilize the platform to achieve a reliable frequency out of the box.

          3. I think if Rockchip thought they had a part that was going to be hitting 3Ghz, they would have been excited enough about it that all the slides would have mentioned it. Also clock speed is not were the excitement is going to be for the next gen mid to high end Rockchip and Amlogic parts.

          4. And if they would’ve mentioned ‘3 GHz’ each and every of their customers would’ve been condemned to deliver 3 GHz 😉

            3 GHz is expensive both with regard to consumption and heat dissipation.

            Currently playing around with Apple’s M1: when throttling kicks in on the passively cooled MacBook Air clockspeeds of the performance cores are reduced by e.g. 17% and consumption drops by 33% at the same time. DVFS at work.

          5. Based on what others have done with A76 on 7nm (RK3588 appears to be 8nm), I am going to revise my expectation:
            2.2GHz A76 + 1.8GHz A55

          6. I’ve been told the 8nm Samsung process used on RK3588 is closer to Intel 10nm process (according to Intel), so 3.0 GHz is indeed not realistic.

          7. Yeah, I forgot this is going to be an “old” part at release. It was supposed to be this year’s release. It will already be well into its original life cycle when released next year. My expectations have been adjusted accordingly.

        4. One of my friend who works in rockchip told me that they have modified the Cortex-A76 in RK3588 and made its IPC higher than X1. However they added some instructions, so you may need to use a modified compiler to get best performance.

          1. sounds a bit strange, it’s not more an A76 if they did such a thing. Also why would they call it A76 if it’s higher ? After all they were calling their A12 “A17”, they would hardly do the opposite.

      1. The slides only mention LP-DDR4x and LP-DDR5. It’s likely that it can still do LP-DDR4, but not regular DDR4, which is quite different.

      2. The reason you can’t use DIMM RAM modules with SOCs is because the DIMM RAM module buses are much wider than most SOCs support. Does the RK3588 support a wide enough bus to allow RAM modules to be used?

  2. this rk3588 and the midrange rk3566 (more info coming soon ?) are my most expected SoCs since the roadmap release
    linux support + a ton of IOs <3

    i can’t wait to see the next-gen pinebook pro & SBCs

        1. It could be. The 8nm process is more expensive, so a low-cost chip like RK3566 will be cheaper that way, and Cortex-A55 cores are already efficient.

  3. So … what CPU performance will this provide? Comparable to Celeron? Pentium? Core i3?

    I’m still very impressed by my Gigabyte Brix: 4-core Celeron, in a nice enclosure, for 149 euro incl 21% VAT. And a Celeron Chromebook costs about 190 Euro

    So … I wonder about the market segment of Rockchip RK3588. I think it’s hard to beat Intel on the low price segment.

    1. Core i3 performance, I think. Raw integer performance should be at the level of an i5-4460. Maybe higher if it clocks north of 2.3 ghz.

      1. I really really doubt it can reach an i5 4th gen by any means. That i5 consumes 84w and is clocked at 3,4Ghz, RK3588 is an embedded SoC ready to very low consumption. In a real world, my guess is that i5 would obliterate RK3588 on daily tasks, specially on stressful ones. Even the GPU will be better, I don’t have any doubt, as it is a desktop CPU. I know it’s really tempting to always compare ARM with X86 CPU’s but we need to be realistic, magic does not exist, and if you have a SoC that consumes 4-6W at 8nm and a CPU that consumes 84W at 22nm, there is no chance. It would be different if that i5 would be an ultra low power one, like 3rd gen i5 3427U, in that case i5 would still be faster, but the margin will be closer.

        In a raw comparison, I would say RK3588 would be like a Pentium J50XX, which is very good anyways.

  4. Yeap, jack of all trades!
    Imagine how long it will take for Linux support ….

    And yes Socketed (not-soldered) RAM !

    4 power + 4efficiant but why there is no 2power + 2 efficient combo – this seems the best to me
    Can you name such SoC from any vendor?

    1. The mediatek has some like that, i think that it was the MT8173

      And Qualcomm made one like this long time ago, the SD820

    2. Soldered takes way less space on board and is much les expensive. Plus I’m not aware of any LP-DDR4 RAM sticks (or at least not any easily available).

      1. ODROID-H2(+) has 2 empty SO-DIMM slots at ~$120. The larger form factor is better IMO, allowing more ports. I’m pretty sure you can get 32 GiB of DDR4 SO-DIMM cheaper than any 32 GiB LPDDR4(X)/5 option for RK3588 is going to cost. About $85 or $100 for 2×16 of the slower DDR4-2666.

        So if RK3588 doesn’t support it, that’s a bit disappointing. It wouldn’t have made sense if max supported memory was 4-8 GiB, but it does at 32 GiB.

        1. You are comparing x86 architecture to ARM64 architecture. Not a valid comparison IMHO. FYI, I do have a H2+ also which I don’t really use much. N2+ and RK3399 all the way for me. H2+ it’s also much larger, requires more Watts and it’s technically a PC. I think you should stick to x86 hardware.

          1. ODROID-H2+ is minuscule compared to typical x86 hardware. There’s nothing stopping anyone from sticking a more efficient ARM SoC in that larger form factor. There is also a trend towards increasing the wattage of ARM CPUs, as seen with the Snapdragon SQ1, the new ARM Cortex-X1 core, and Apple M1.

            ARM + SO-DIMM would allow a user to buy exactly the amount of RAM that they want, for less money. If it uses more power than LPDDR, keep it plugged into a wall instead of a battery.

          2. What he’s explaining to you is that the real benefit of ARM vs x86 on SBCs is the cost+consumption for a given performance level. If you start to significantly inflate the price and development costs by making it much more extensible, you will not benefit from the price difference (the ARM will be more expensive due to lower volume and higher development cost) and the performance will be lower. ARM can be interesting on the high end (servers) when the development costs are largely offset by the power bill and you can afford to buy a twice more expensive machine for the same performance.

          3. It is surprising Rockchip have not dropped the node lower and tweaked the RK3399 / pro design, as it has had a long life and software, hardware support.

          4. On the other hand, if it continues to sell as-is, they would be stupid to change anything there!

          5. So you would not like a cooler running SoC with a higher clock at a node size that gives a higher yield from silicon for Rockchip.

  5. What do you think the RK3588 will cost? $30? $35?
    I wish they’d drop the price on the RK1808 down to $5-6. At $10 it is too expensive for our use case.

  6. Now that made me drool. I wished I could already have one in my hands.
    Yesterday I reviewed a AWS server with 64-cores N1. A derivative of the A76. All clocked at 2.5Ghz while performing like an A73 that would have to be clocked to 3Ghz.
    I hope the A76 will be clocked close to 3Ghz. The A55 at around 2/2.5Ghz.
    The max ram of 32 is amazing. I would want 16GB. With N2+ I don’t have enough with 4GB +2GB zram + 4GB swap. So 16GB will do.
    Too bad the human malware has slower the release so much. I want it so badly. Can’t wait to see how it does with my tasks.

  7. I would expect that one of the PCIe interfaces is meant to connect to an on-board NVMe device. At the moment, NVMe chips in BGA package are about twice the price of eMMC/UFS for the same capacity, but also faster and more durable, so I’d always pick NVMe when given the choice, in particular when the other components are expensive as well.

    For low-cost boards, eMMC or (shudder) SD is the only choice anyway, as UFS and NVMe only start in relatively large capacities of 32GB and 64GB, respectively.

    1. Fairly sure that answer is “never”. SBSA doesn’t really make sense for embedded SoCs, since you need a ton of device drivers for all the integrated peripherals anyway. There is really not that much difference in what UART or watchdog driver you use, or whether the PCI controller is compliant when you need platform specific kernel support to get the GPU and display up.

  8. I’m also impatient to get one, but I already predict that once the first SBCs are out, half o f the readers here (me included) will say “oh shit, they put the SoC on the wrong side” (regardless of the side), and “what, I need a FAN on my SBC now?”. Let’s keep in mind we’re talking about a quad-A76 (i.e. a mini-xeon) where thermal design will certainly be a challenge.

    1. *stares at phone* Uhh, it doesn’t have a fan, either. But I’m not going to be one of the people who complain about fan. I will complain if they don’t put the SoC on the bottom so it can be heatsunk to the case properly. I’ll also complain about power delivery is it uses any kind of USB for that purpose.

      Mr. Cubie, please use a 5.5mm barrel with a compliance range of 7.5V to 19V as a minimum. More range on either end is welcome as well. If you’re putting SATA on it and need a 12V output (in case someone wants to waste a SATA port to hook spinning rust), then a lower input range of 12V would be acceptable. That’s not my use case, but I understand that there are many here (Thomas probably) who will desire it–SATA ports that is).

      To be selfish, I would just want a PCI-E 3.0×4 NVME port, two GigE ports (or faster?) and as much and as fast USB as you can give me without resorting to a hub. I’ll buy one. 🙂 Clearly a huge market.

      1. > I understand that there are many here who will desire … SATA ports

        Just why? To attach large spinning rust there exist less expensive ARM boards like e.g. an ODROID HC2 or HC4 (in my country I need to multiply Radxa $ prices by 1.3 to get amount of € I’ll really pay so I would expect an RockPi 5 with minimum DRAM without enclosure and PSU to start at 130€. Wasting this setup for a pile of spinning rust would be a bit moronic).

        Since PCIe Gen2 lanes and SATA are multiplexed exposing them in a mPCIe/mSATA slot would’ve been a smart idea in the past (just like Clearfog boards did). But by the time the first RK3588 SBC will be available and software support in shape this interface connector is really too outdated to be of any use.

        1. Why SATA? I’m always looking for a good low-cost NVR machine. My ideal has fast H.264/H.265 decoding and image processing, a good NPU (for object detection), SSD (for the database), SATA (for terabytes of video on cheap spinning rust), and two NICs (one for the camera network segment, one for the DMZ network segment). This SoC is close to perfect. If it didn’t have the SATA, I’d have to use a USB bridge, which is an added cost, and a source of flakiness, particularly in UAS mode under Linux.

          1. Having the hardware is step 1. When do you think software will be ready? You’re talking about something usable with Linux and not just an Android black box, aren’t you?

          2. Hardware for coming sbc (may it be arm32?, arm64, x64, risc64, …, ram options, storage solutions, networking speeds, io, multimedia) is probably no more limitation to most average users for desktop-like communication, office work or entertainment (on average levels for image processing).
            New limitations come from e. g. admins, that are forced to build software related workflow “interrupts” for users (web browsing on granting cookies or data transfer permissions to data dealing companies as one obvious ‘increasing’ occurance) for ensuring their livings income?
            Hopefully ai can be of help with that?
            If memory bandwidth from ram or ssd storage achieves rates like 2-5GB/s on 100-200$ devices, having user-side endurances like 3-5 years of usability (including software compatibility for programs, e.g. x64 to arm64 including purchase for updates or developers going ‘MultiArchitecture Binaries’?) who thinks about updating then to lp-ddr6, without need, or usb4 anymore, besides with getting new devices?
            rk3588 will be one of last devices, that are recognized apart from desktop line/laptop computers or phones, as a “sbc”, with e.g. external single supply voltage (vs. ATX 3.3,5,12V, USB-PD5,9,15,20V, PCIe3.3,12V)?

          3. (USB-PD5,9, ***12*** ,15,20V, SATA(3.3),5,12V)

            SATA stays on 6Gbit/s since around mid of 2008. Later revisions from 2013, 3.2 (-3.3), introduced SATA Express (what is in reality PCIe protocol and additional SATA backward compatibility) 8Gbit/s and 16Gbit/s, but no devices are available nor requested, since (consumer side) rise of NVM-E(xpress protocol) on M.2 connectors or PCIe.

          4. I’m developing the software myself: Moonfire NVR. Open source, runs on Linux. It’s a work in progress that currently has solid recording, a crude prototype UI, no ML-based video analytics yet. I slowly improve it when I find time.

          5. > runs on Linux

            So you need hardware supporting Linux and not just Android (some relevant stuff locked down in BLOBs and the user having to remain on an ancient kernel version forever). Nobody knows today whether Rockchip will go the extra mile and care about Linux support for RK3358…

          6. Oh, I misunderstood your “when do you think software will be ready?” earlier. But they did pretty well with rk3399, didn’t they? I don’t know how long it took, but now it seems like a pretty well-supported chip used by a bunch of SBCs.

          7. I don’t think they’ll keep a dual-GigE chip for android only. They probably figured there is a market for home and small office appliances.

          8. > I don’t think they’ll keep a dual-GigE chip for android only.

            Well, for me adding legacy SATA is even more an indication of this design not being ‘Android only’. But only time will tell. Let’s talk again in 2022 or once RK3566 is there (since also equipped with SATA, PCIe and 2x RGMII interfaces with TSO).

          9. > RK3566 … equipped with SATA, PCIe and 2x RGMII interfaces with TSO

            Silly me, I forgot that we have more recent information about RK3566 in the meantime. According to https://www.cnx-software.com/2020/02/01/rockchip-rk3566-rk3588-rv1109-socs-coming-in-2020-based-on-rockchip-processor-roadmap/#comment-575631 SATA and 2nd GMAC are gone.

            The whole blog post is also interesting since showing that RK3588 was initially announced to be available Q1/2020, then ‘Q3/Q4 2020’ and this times they’re telling ‘Q2/Q3 2021’… and still commenters get excited about each delay announcement.

          10. I’m trying to find more details about RK3566 and RK3568 (PCIe 3.0), as it looks like they’ve changed the specs again.

          11. I’m trying to get documents because Rockchip talked about those two processors plus others for NVR and IP cameras at the RKDC.

          12. I got confirmation from RKDC about 22nm process. The VAMRS post does not mention RGMII + 2 SERDES so I’m not sure it’s still correct. RK3566 does have PCIe and SATA.

        2. i think m.2 is answer for such purpose, you can also expose sata as well as pcie and all within standard ( x2pcie and sata is possible on m.2 key B, and wiki say that even with key M, and there is nothing holding back with providing pcie x1 interface electrycaly). additionally it’s also possible that this interfaces will be used to provide more usb3 ports (as in rpi4 in example) and not exposed, which i think is fair enaugh and will enable board designers for more flexibility.

          1. > x2pcie and sata is possible on m.2 key B

            And then there’s only one single PCIe device available which absolutely makes no sense at all (the strange single lane SATA adapter Firefly guys designed for their RK3399 board years ago where they chose Key B for PCIe).

            For SSDs key M would be ideal since vast majority of M.2 SATA SSDs is coded B+M. But users might want to do other stuff with a Gen2 PCIe lane and then the slot in question would’ve be better key A or E.

            And right now nobody knows whether the individual PCIe lanes on this SoC support x1/x2 configurations too or it’s x4 as with RK3399 so without an additional PCIe switch adding to BOM costs PCIe situation is ‘all or nothing’ and you waste up to 3 lanes when you connect a single lane controller via PCIe.

          2. I was informed that PCI-e 3.0 supports up to 4 lanes, and can be configured as: x4, x2 + x2, x1 + x1 + x1 + x1

          3. USB highly dynamic device changes (1-3m, most distant hub connections ever? 5hubs2.0-127hubs3.x?)
            PCIe(2.x-4.0)/SDhx.eMMC static onboard setups
            S.ATA static distant (1-(2)m) setups
            Bluetooth5.2(2Mbit/s)/802.11ax(75–1200 MB/s)(/GBLan) increasing user acceptance for mass data transfer, networking and synchronisation

            sbc == versatile ?

          4. correction on SDxc/SDuc card slots: 3000-5000 (-10000) insert&eject cycles (data reliability guaranteed for a maybe ~10 years on standardized surroundings?)

  9. I hope to be able to use this board as an ESXi ARM server next year. My needs seem to be a little special as I want to create 10-12 Android VM running inside and give access through the net as “anywhere” virtual android devices. I’ve had that idea since years ago and now with that board it seems quite possible, unless there is no support for any virtualization appliance 🙁

      1. Maybe you know something I don’t. Are you talking about a specific vendor or software? or about VMOS?, please do share your knowledge 🙂

    1. It’s Rockhip’s fault for not having proposed any updated high-perf chip for a while. Most of their recent chips were entry level A7 or A53 that are very likely much cheaper from other vendors, and the only high-end chip they produced since RK3288 was RK3399 which only has two powerful cores and which used to be very expensive. ASUS has a tablet using it. But since then we’ve seen a lot of competition in this area, with$150 tablets using for example the much faster Helio X27 (2×2.6GHz A72 + 4×2.0 A53 vs 2×1.8+4×1.4 for RK3399), coming in a smaller thermal envelope. And you even find the still faster Helio X20 at 2.3 GHz in $60 tablets. So what remains as a differentiator for the RK3399 is not suitable to such use cases (PCIe, GbE, …). Let’s face it, RK3399 is *old* and until RK3588 is there, Rockchip has nothing else to offer to this market.

    1. “Odin” is upcoming top-performance Mali Valhall GPU. Previous post about RK3588 here mentioned “Natt”, which is also Valhall, but optimised for efficiency. Looks like Rockchip is aiming for the moon with this chip.

    1. 4x Cortex-A76 and 4x Cortex-A55 on Samsung 8nm can’t be that hot.

      The real question, will it ever make it out of the oven? I’ll just get an AMD APU instead.

      1. > The real question, will it ever make it out of the oven

        And will anyone still care ? Especially when you consider how long it usually takes before their chips are finally supported in mainline (with missing features like ddr_freq) and the extreme pain it is to work on the kernel on rockchip devices, by the time it starts to become usable the smallest Atom will be cheaper, faster and less power-hungry, as often :-/

Leave a Reply

Your email address will not be published.

Advertisement
Advertisement