Last May, we wrote about RK1808 AI Compute Stick, a USB stick with Rockchip RK1808 dual-core Cortex-A35 processor also featuring a 3.0 TOPS neural processing unit to accelerate AI workloads at low power.
As I understood it, it was available for purchase, but you had to contact a Rockchip FAE by email in order to get one. Now, you can easily buy one online by getting the Toybrick TB-RK1808 AI Compute Stick on Seeed Studio for $86. Just ignore the “Core i3” in the title, we’ll see why it’s there further below.
TB-RK1808 AI Compute Stick specifications:
- SoC – Rockchip RK1808 dual-core Cortex-A35 processor with NPU
- AI inference performance – 3 TOPS for INT8, 300 GOPS for INT16, 100 GOPS for FP16
- System Memory – 1GB LPDDR
- Storage – 8GB eMMC flash
- Host Interface – USB 3.0 type-A port
- Power Supply – Via USB port
- Dimensions – 82 x 31 x 13 mm
Despite the company claiming minimal requirements are an Intel Core i3 Ubuntu 16.04 computer with a USB 3.0 port, 2GB RAM and 4GB free storage, development is said to be possible on Windows, Linux, macOS, and Arm Linux (e.g. Raspberry Pi). I suppose what they mean is that developing and training models requires a more powerful computer running Ubuntu, while you can use lower-end hardware to run the model on the compute stick as is normally the case with this kind of hardware e.g. Intel Neural Compute Stick 2.
The user manual shows documentation and tools are stored on the USB stick itself.
You’ll find some of it on RKNN-Toolkit repository on Github, but documentation specific to RK1808 AI Compute stick such as “RK1808_AI_Compute_Stick_User_manual_EN.pdf” or “RK1808_RKNN_SDK_DEVELOPER_GUIDE_EN.pdf” does not seem to be available publicly. There’s however some documentation on ToyBrick’s Wiki that includes demos for Mobilenet and Yolov3.
The price tag ($86) is a bit higher than I would have expected, as Orange Pi AI Stick 2801 with similar performance previously sold for $69. It did not include PLAI software tool (~$150 extra) however, which makes RK1808 AI Compute Stick quite cheaper since AFAIK it does not require the purchase of software tools.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
ROCK Pi NN HAT, rk1808 inside, designed for Raspberry Pi 4 and ROCK Pi 4, 25$
That’s more like it. But when will it become available?
We will announce it soon.
Could you please detail working temperature range? Edge computing could be put outdoor and I’d need industrial range
The host code for when NPU stick in used in an ARM host is broken. It is assuming the ARM host is a RK3399 and then it looks for the internal RK1808. Bug report is here, hoping Rockchip hurries up and fixes it.
https://github.com/rockchip-linux/rknn-toolkit/issues/2
The host code works properly on x86 and Mac.