Rockchip RK3399Pro EVB and NPU Performance Demos (Video)

Rockchip RK3399Pro was announced as an updated version of RK3399 processor with an NPU (Neural Processing Unit) capable of delivering 2.4 TOPS for faster A.I. workloads such as face or object recognition.

There haf been some delays in the past because of a redesign of the processor that placed the NPU’s RAM on the PCB instead of on-chip for cost reasons. Eventually we got more details about RK3399Pro, and today I also received a 15-page presentation with some more information about the software, and processor itself.

Rockchip RK3399Pro EVB
Rockchip RK3399Pro EVB – Click to Enlarge

But even more interesting, that’s the first time I see Rockchip’s official RK3399Pro EVB (Evaluation Board), and the guys at Khadas uploaded a video to explain a bit more about the board, and showcase the NPU performance measured up to 3.0 TOPS with an object recognition demo, and an “body feature” detection demo – for the lack of a better word – running in Android 8.0.

Eventually, Shenzhen Wesion (Khadas) will launch their own RK3399Pro board with Khadas Edge-1S now “available” on Indiegogo.

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

ROCK 5 ITX RK3588 mini-ITX motherboard
Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
9 Comments
oldest
newest
willy
willy
6 years ago

This demo is quite convincing (well done Gouwa). I think there’s more than enough performance to detect moving pedestrian for car automation. It looks like hardware made a lot of progress over the last few years, I’m even wondering if the new NPUs will not quickly make GPUs obsolete for general purpose computing (they’ll become graphics chips again).

blu
blu
6 years ago

NPUs are matrix multipliers. Any potent matrix multiplier (case in point — GPUs) can be a successful inference machine, but matrix multipliers going gen-purpose is about what _else_ they can do well, aside from multiplying matrices. In this aspect GPUs are, and will likely remain, the much better massively-parallel gen-purpose machines for the foreseeable future.

geokon
6 years ago

I think the elephant in the room is that these NPUs don’t seem to support OpenCL or anything similar.. and who wants to be stuck programming this stuff? You’re gunna have to re-invent the wheel and rewrite your whole processing stack for this one custom chip. Doesn’t sound fun

I keep seeing the same pattern, you just get little demos from the chip maker and then they’re expecting device companies to do the heavy lifting. I’m no expert, but to me it seems to only make sense in price-sensitive mass produced devices (something like Alexa or a smart security cam)

blu
blu
6 years ago

Absolutely. APIs and programming models is another side to all this.

crashoverride
crashoverride
6 years ago

NPUs typically do not have the precision required to be useful as a general purpose processor. The document reference in the article for the NPU used in RK3399 Pro states “1920 INT8 MACs, 192 INT16 MACs, 64 FP16 MACs”. The focus is on integer (8/16 bit) calculations and “half” precision (16bit) floating point.

willy
willy
6 years ago

With 1920 int8 MACs you can do awesome stuff 🙂 I’m starting to think there may be a way to use this for pattern matching for new generations of compression algorithms. I hope (and expect) they support saturated arithmetic by the way (I haven’t checked the doc yet).

e97
e97
5 years ago

INT8 is used for inference. INT16 can be used for training if you know what you’re doing and the domain can handle the precision. There’s an Alphabet presentation that goes into detail on this topic.

Lalith
Lalith
6 years ago

24 fps detection … very nice. Thank you for posting this video

Gouwa
6 years ago

There is a comparison between RK3399Pro and other NPUs, you may confused about the other NPU models, more details below:
– T: NVIDIA TX2
– H: HiSilicon Kirin 970
– A: Apple A11
– M: Movidius Myriad X

Have fun in the AI world!

Boardcon Rockchip and Allwinner SoM and SBC products