Last year, Movidius announced the Fathom compute stick, a USB stick containing a Myriad 2 vision processing unit (VPU) used to accelerate machine learning applications. Today, Movidius announced an updated version called the “Neural Compute Stick,” which has better performance and machine learning framework support, as well as a cheaper price as a result of gaining access to Intel’s large manufacturing facilities.
Offline Deep Learning Accelerator
The Movidius Neural Compute Stick is the first “self-contained” AI accelerator in a USB stick format. It’s targeted at product developers, researchers, and makers who want to reduce their research and development costs when testing their machine learning applications.
According to Movidius, now an Intel company, the Neural Compute Stick customers can use either Intel’s Nervana cloud to train their neural network models or any other training setup. The compute stick will run the models regardless of where they were trained.
As the Neural Compute Stick is an inference accelerator, it doesn’t require an internet connection to function. This brings several benefits, such as better privacy, as the data on which the neural networks have to work can remain on a local device. The inference operations are also more efficient because they happen in real-time, as opposed to across the world in a data-center, and there is no internet connection lag or possibility of service disruption.
Software Capabilities And Performance
Developers will be able to compile a trained Caffe-based convolutional neural network (CNN) into an embedded neural network that is optimized to run on the Myriad 2 VPU inside the Neural Compute Stick. Myriad 2 promises around 100 GFLOPS using only 1W of power and about 10-15 inferences per second.
The stick also supports layer by layer performance metrics for both industry standard and custom-designed neural networks to enable tuning for low-power real-world performance. Where one Neural Compute Stick is not enough, the developers will also be able to use up to four compute sticks at once to drastically increase inference performance for their applications.
The Movidius stick also supports multi-stage multi-task convolutional neural networks (MTCNN), which can be used to obtain improved facial recognition. Intel will talk more about these capabilities at the July 22-25 conference on Computer Vision and Pattern Recognition in Honolulu, Hawaii.
The Movidius Neural Compute Stick is now available for purchase through select distributors for only $79. The stick will also be available for purchase at the conference in Honolulu.