Toybrick RK3399Pro Board is equipped with Rockchip RK3399Pro AI processor

We’ve heard about Rockchip RK3399Pro processor that adds an NPU (Neural-network Processing Unit) to Rockchip RK3399 hexa-core processor for about a year now. We’ve seen the official Rockchip RK3399Pro EVB in action last November, but this type of board is not usually for sale to the general public.

The first RK3399Pro board that you could technically spend money on was Khadas Edge-1S via a crowdfunding campaign, but it’s still unclear when it will ship. Another option that I just discover is Toybrick RK3399Pro (TB-RK3399Pro) single board computer by VAMRS.

Toybrick RK3399ProToybrick RK3399Pro board preliminary specifications:

  • SoC – Rochchip RK3399Pro hexa-core big.LITTLE processor with 2x Cortex A72 cores up to 1.8/2.0 GHz, 4x Cortex A53 cores @ 1.4 GHz, and an Arm Mali-T860 MP4 GPU with OpenGL ES 1.1 to 3.2, OpenVG1.1, OpenCL 1.2 and DX 11 support, and NPU delivering up to 3 TOPS
  • System Memory – 3 or 6 GB memory
  • Storage – micro SD card slot + 16 or 32 GB flash
  • Video Output & Display I/F
    • 1x HDMI 2.1 (TBC)
    • 1x DisplayPort 1.2
    • 1x eDP connector
    • 1x MPI DSI interface
  • Camera – Dual MIPI CSI and dual ISP up to 13MP
  • Audio
    • 3.5mm headphone jack
    • Speaker header
    • Built-in microphone
    • Support for 8-channel digital microphone array input;
  • Connectivity – Gigabit Ethernet, WiFi module
  • Expansion
    • 4-lane PCIe connector
    • mPCIe socket
    • I/O expansion header
    • RS232 header
  • Debugging – 1x micro USB port for access to the serial console
  • Misc – Power, reset, and “other” buttons; fan header
  • Power Supply – Via USB type-C port
  • Dimensions & Weight

TB-RK3399Pro board comes pre-installed with both Android and Linux operating system in dual boot configuration. Software to leverage the NPU is also included with support for TensorFlow/TensorFlow Lite/Caffe inference models.

Toybrick RK3399Pro SBC is up for “pre-order”, but in the sense that they check people’s interest first, asking people to fill in some information about their project before pre-order. The price is still mostly unknown as the company only mentions the price will be below. The announcement was written on December 11, 2018, and at the time the first samples of the board were supposed to ship in mid-January.

Via miniNodes

Support CNX Software - Donate via PayPal or become a Patron on Patreon

14
Leave a Reply

avatar
5 Comment threads
9 Thread replies
0 Followers
 
Most reacted comment
Hottest comment thread
9 Comment authors
goussackVincetkaisertheyguyukcnxsoft Recent comment authors
  Subscribe  
newest oldest most voted
Notify of
Johannes
Guest
Johannes

Cool! The board is available on taobao here
【TB-RK3399Pro 开发板 6核64位AI开发板 Linux+Android双系统】https://m.tb.cn/h.3tLmHhG?sm=d15228 点击链接,再选择浏览器咑閞;或復·制这段描述¥BknzbIrLKkR¥后到👉淘♂寳♀👈
1499 rmb for 3g/16g, 1899 rmb for 6g/32g. Rather expensive (220usd / 280usd), but maybe the start of the competition where more boards will follow. I guess 6g is rather important because the ram will be shared also for the machine learning inference.

blu
Guest
blu

AFAIK ram is not shared between the cores and NPU, so it’s likely a 4 + 2 config.

Johannes
Guest
Johannes

Yeah sorry that’s what I meant, that part of the RAM is most likely dedicated to the NPU, but I suppose exclusively. So therefore 3GB is more like the 2GB entry level Rockchip boards, with 1GB extra for the NPU. The Intel Movidius has 512 MB, and I guess for a reduced model size that runs on this device in decent performance that should be enough.

willy
Guest
willy

The design looks pretty nice and well thought. There’s a real PCIe slot. And there’s even a RESET button! I expect that it will be too expensive for most of us though.

Anton Fosselius
Member

If we ignore power efficiency and size, how does it compare in $/performance to the X86 offerings? Or Nvidias boards?

For me, performance must be really good to motivate that price..

blu
Guest
blu

In that price range ($280-300) you can already get a macchiatobin mITX featuring 4x [email protected], but sans ram.

Johannes
Guest
Johannes

It is rather expensive, you can get a Intel J5005 NUC with 4G ram and a 128G m2 ssd (for less than the 6G version . This would have superior CPU performance, but not the AI accelerator. If you buy an additional Neural compute stick, you’d have better performance, but also higher price. But keep in mind that Intel CPUs are quite expensive now, especially the Atoms seem to be in low supply and therefore artifically expensive.

theyguyuk
Guest
theyguyuk

And a higher electric bill with intel

tkaiser
Guest
tkaiser

> Intel J5005 … superior CPU performance

Not that much, such a J5005 is roughly 40% faster with single-threaded loads compared to an A72 at 2GHz in RK3399 and with multi-threaded loads utilizing all cores the difference is even slightly less.

tkaiser
Guest
tkaiser

> Intel CPUs are quite expensive now, especially the Atoms seem to be in low supply and therefore artifically expensive

I’ve no problem getting an ASRock J4105B-ITX for 75€ ($85) while the same board with a J5005 (then named ‘J5005-ITX’ of course) costs 119€ ($135). The $50 difference is J5005 vs. J4105 and since the latter is only 10% slower it’s an easy choice IMO.

Vince
Guest

Lord…..who makes a tiny board that expensive…..when there are other quad core boards at just 10 dollars…..ridiculous.

goussack
Guest
goussack

can modern graphics cards be launched on this board?
russians launched various video cards on their board with baikal (mips arch) and with elbrus (e2k arch)
as I know the main problem is the presence of x86-code or uefi-byte-code in “firmware”