Get deep learning and AI processing in your pocket for $79 with Intel's Movidius Neural Compute Stick
Intel's subsidiary Movidius has launched a pocket AI accelerator, the Neural Compute Stick (NCS) that enables host devices to process deep neural networks natively (i.e. at the edge). This enables AI researches to optimize offline AI applications.
The NCS plugs into the USB 3.0 port of the host and is powered by the Myriad 2 Vision Processing Unit (VPU) that can reach 100 GFlops of computing performance with a power consumption of just 1W. The VPU allows translation of a trained Caffe-based conventional neural network (CNN) into an embedded neural network and its execution. The Myriad 2 VPU is complemented with 4GB of LPDDR3 on-chip memory. This allows the NCS to offer native FP16 precision with over 10 inferences per second. Movidius claims that the new 'Acceleration Mode' and 'Multi-Stick Mode' help in linear performance increases and the ability to combine multiple sticks to offload work from the CPU.
The NCS is available for US$79. Such a low cost of entry could help AI researchers adopt low cost, low power but capable inference processors that provide privacy in processing AI data without the hassles of server latency and high power requirements of high-performance CPUs. Intel will be demoing a workflow implementing the Caffe/NCS system at the Computer Vision and Pattern Recognition (CVPR) conference in Honolulu, Hawaii. The conference is scheduled between July 21 - July 26, 2017.