New boards will accelerate AI-related tasks and boast specs more akin to a desktop PC.
Building a slide deck, pitch, or presentation? Here are the big takeaways:
- Single-board computers will be released with a dedicated chip for accelerating AI-related tasks.
- The boards are aimed at developers who want powerful low-cost machines for testing out AI, IoT, automotive and other applications.
Single-board computers have traditionally been relatively underpowered, but 2018 will see the release of boards with serious specs — particularly when it comes to running AI-related tasks.
The first, the HiKey 970, is an octa-core board that uses the same HiSilicon Kirin 970 SoC [system-on-a-chip] found in the Huawei Mate 10 phone, and which is targeted at developers working on everything from robotics to autonomous vehicles.
The board will pack 6GB of DDR4 RAM, can support 4K displays via HDMI 2.0a, has USB 3.0 with type-C support and runs Ubuntu, Debian and Android.
The SoC features a dedicated Neural Processing Unit — a chip designed to carry out AI-related tasks, such as image recognition, more rapidly.
SEE: Raspberry Pi 3 Model B+: An insider's guide (free PDF) (TechRepublic)
In tests, the NPU on the Kirin 970 SoC was found to be orders of magnitude more efficient at running pre-trained Convolutional Neural Networks (CNN) to recognize objects in images, compared to the SoC CPU found in the Google Pixel 2 and the CPU on the earlier Kirin 960 SoC.
In practice this means the board can be used to carry out more complex AI-assisted tasks in closer to real time. On the Mate 10 the NPU is used to dynamically optimize the camera settings when taking photos, to accelerate offline translation, and to classify photos.
"The Hikey 970 enables new products in industry, audio devices, robotics, audio devices, video devices, automotive and we can't wait to see what people do with this product," said Linaro CEO George Grey at the Linaro Connect Hong Kong 2018 conference this week.
The board is expected to be available shortly but has not been priced as of yet, although is likely to be pricey, with the 970's predecessor the HiKey 960 selling for $239.
- System-on-a-chip: Kirin 970
- CPU: 4 x Cortex A73 @ 2.36GHz, 4 x Cortex A53 @ 1.8GHz
- GPU: 12-core Arm Mali G72-MP12 GPU, NPU with 256MAC/cycle @ 960MHz
- Memory: 6GB LPDDR4
- Storage: micro SD card slot
- Video Output: HDMI 2.0a
- Camera: 4 lanes CSI + 2 lanes CSI
- Connectivity: Gigabit Ethernet, wireless module
- USB: 2 x 3.0, 2x type-C
- Expansion: mPCIe connector, 40-pin 96Boards low-speed expansion header, 60-pin 96Boards high-speed expansion header
- Power: 8 to 18V via power barrel
ROCK960 Enterprise Edition
Also announced was the Rock960 Enterprise Edition, a board with specs more akin to desktop PC, for those interested in developing applications using 3D graphics and machine-learning.
The Rock960 Enterprise Edition is based on a hexa-core processor, has a powerful GPU, and a dedicated Neural Processing Unit capable of 2.4 TOPS (trillion operations per second). In tests this NPU can accelerate the performance of convolutional neural networks —such as Inception V3, ResNet-34 and VGG16—by almost 100x, with some benchmarks showing it beating both the Apple A11 chip and Kirin 970 SoC.
As with the HiKey 970, these figures reflect the performance when running pre-trained machine-learning models, rather than when training models, a task that typically requires even greater parallel processing power.
The board also sports a massive 3 x USB 3.0 and 5 x USB 2.0 ports, dual SATA 3.0 ports, up to 8GB RAM and powerful add-on boards via its PCIe x16 slot, however its worth bearing in mind that the SoC itself only supports 4 PCIe lanes.
Again, given the specs, expect the board to cost far more than $35 Raspberry Pi 3 Model B+ when it launches in Q2 of this year.
- System-on-a-chip: Rockchip RK3399Pro
- CPU: Hexa-core processor with 2 x Cortex A72 cores, 4 x Cortex A53 cores and NPU capable of 2.4 TOPS
- GPU: Mali-T860MP4 GPU OpenGL ES 1.1 /2.0 /3.0 /3.1 /3.2, Vulkan 1.0, Open CL 1.1 1.2, DX11
- Memory: Up to 4GB RAM
- Storage: Dual SATA 3.0 port with RAID 0/1 support
- Video: HDMI 2.0/eDP up to 4K @ 60 Hz
- Camera: Dual MIPI CSI interfaces
- Connectivity: Gigabit Ethernet, 802.11ac WiFi
- USB: 3 x USB 3.0 ports, 5 x USB 2.0 ports
- Expansion: PCIe 2.1 x16 slot
- Remote Management: 4G LTE, management port
Grey also revealed the Ultra96, Xilinx Zynq UltraScale+ MPSoC board, which is targeted at machine learning, automotive and IoT, and due to be shipped in about six weeks time.
More information about all of the boards can be found at this site.