Rockchip RK3399Pro was expected to be the first processor from the company featuring an NPU (Neural Processing Unit) to accelerate artificial intelligence workloads, but eventually went through a redesign, and the community found it gave birth to RK1808 NPU by looking at a device tree file in the Linux kernel.
Rockchip has now made the NPU official at CES 2019, and we now know a little bit more.
Here are the specifications shared by the company:
- Dual-core Arm Cortex-A35 CPU
- NPU computing performance up to 3.0TOPs supporting INT8/INT16/FP16 hybrid operation
- VPU supporting 1080P video codec
- Built-in 2MB system-level SRAM
- Display – MIPI/RGB video output
- Camera – MIPI/CIF/BT1120 Camera video signal input with built-in ISP
- Audio
- Microphone array support with hardware VAD function for low-power monitoring and far-field wake-up
- Audio output
- I/O
- PWM/I2C/SPI /UART
- USB3.0/USB2.0
- PCIe interface
- Support for Gigabit Ethernet and external WiFi/Bluetooth modules
- Manufacturing – 22nm FD-SOI process
It shares many similar functions as standard media SoCs, but it appears to lack a GPU for example, it’s not a platform capable of handling advanced user interface with 3D graphics.
The NPU is compatible with TensorFlow, MXNet, PyTorch and Caffe frameworks, and the company provides a Linux system AI application development SDK supporting C/C++ and Python programming languages.
Rockchip RK1808 solution is expected to end up in various AI applications in the domain of security, education, automobile, smart wearable devices, home appliances, storage, etc.. thanks to AI functions including voice wake-up, speech recognition, face detection and attribute analysis, face recognition, posture analysis, target detection and recognition, image processing, an so on.
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
So this is the same NPU as on the (long announced) RK3399Pro, which we’re still waiting for. Apparently coupled with the two A35 cores to drive the NPUs, but without GPU, so targeted at non-display embedded devices. I wish instead of announcing more devices, Rockchip would finally release the original RK3399Pro. The NPU was announced, if I remember correctly, on last year’s CES, so one year without any visible product. The RK3399 is from 2016 and in 2019, we’re not really seeing a new product on that front either. A bit disappointing.
Despite all the billions plowed into Chinese semiconductor industry, what is actually coming out of it? Does any of the hyped AI startups have a chip released yet? I’m aware of Cambricon who supply a NPU to Huawei smartphones, but that’s about it as far as I know.
With enough RAM this might be up to training not too large deep learning models…
Can the NPU use the ARM cores’ DDR5? Any idea what kind of memory bandwidth is possible? (Apparently this has a more important influence on the speed at which deep models can be trained than TOPS/GOPS/GFLOPS etc)
Get a good Nvidia card and train on desktop. Probably 100x faster. Just use this for running the finished networks.