Intel on Wednesday is rolling out the Neural Compute Stick (NCS) 2, the second iteration of its popular self-contained AI accelerator. With this version, the NCS offers better performance, as well as scalability and a path to production.
The NCS comes in a USB format, making it easy to compile, tune and accelerate neural networks at the edge. No cloud connectivity is needed for its on-board processing. Developers and researchers can use it to seamlessly convert and deploy PC-trained computer vision models to a wide range of devices natively.
The NCS 2 is powered by the latest generation of Intel’s vision processing unit (VPU), the Movidius Myriad X VPU, providing a performance improvement of up to 8X over the first NCS.
It’s also supported by Intel’s OpenVINO toolkit, which enables developers to build and train AI models in the cloud and deploy them across a broad range of products. This gives developers wide framework and network support, and even the flexibility to move models to more VPUs, a CPU, integrated graphics or FPGAs.
Meanwhile, with the Intel AI: In Production ecosystem, developers can port their Intel NCS 2 prototypes into production.
When Intel launched the first-generation NCS last year, it quickly sold out, Steen Graham, GM of IoT channels & ecosystem at Intel, told ZDNet.
“We’ve really had very little insight about what the demand was,” Graham said. “A lot of people train models in the cloud, a lot of people do inference in the cloud — but deploying deep learning or AI at the edge, we didn’t know what the stage was in the deep learning community, what their interest was.”
The popularity of the device aligns with trends bringing workloads to the edge, Graham said. According to IDC, 45 percent of data will be stored, analyzed and acted on at the edge by 2019. Intel’s research, Graham said, says that by 2023, over 40 percent of AI tasks will take place on edge devices.
Since last year, tens of thousands of developers have deployed the NCS in a wide range of use cases. For instance, the sensor company Flir Systems recently launched the Firefly camera, a small piece of hardware that uses Intel’s Movidius chips for on-camera inference. They used the NCS Neural Compute SDK to prototype the device.
Intel is highlighting ways developers have used the NCS to advance noble causes, such as detecting contaminated water in developing countries. Developer Peter Ma, who’s backed by Intel, created a system called CleanWater AI to identify water bacteria using pattern recognition and machine learning. The system works with a digital microscope connected to a laptop running Ubuntu and the NCS. Contaminated sites can then be marked on a map.