Intel
Artificial intelligence in USB format
Intel presents a USB stick that accelerates deep learning algorithms with the help of a VLIW processor. It processes the data locally in real time and is therefore particularly suitable for systems without a connection to the cloud.
With the 'Movidius Neural Compute Stick', Intel is launching a USB stick on the market that contains computing resources for artificial intelligence in the form of a USB stick. The Neural Compute Stick contains a Myriad 2 processor from Movidius, a company that Intel acquired in September 2016. Movidius developed the Myriad-2 processor as an image processing chip that can not only process 2D image data, but can also "see" in three dimensions. Intel's original plan may have been to further develop this processor for applications in the field of autonomous driving. However, these plans are likely to have come to an end with the acquisition of Mobileye, as Intel has announced that it will now bundle all ADAS and autonomous vehicle activities under the Mobileye umbrella. Mobileye's EyeQ processor is already being used in more than 300 different vehicles.
Proven with drones
The Movidius platform has proven itself in another market that is also experiencing strong growth: drones. Movidius' most important design win is DJI, the Chinese market leader in the field of multicopters. With the right software, the Myriad 2 processor is able to recognize obstacles in the environment, record 360-degree panoramic videos, take HDR shots and recognize objects.
The Myriad 2 is optimized for mobile applications and devices that are powered by a battery. Its learning ability and intelligence is based on the calculation of neural networks. This requires many small computing units connected in parallel. Intel speaks of 100 gigaflops of computing power with a power budget of one watt. With this computing power, the processor can calculate locally and react in real time without having to resort to resources in the cloud.
The Myriad 2 processor contains a dozen SHAVE cores (streaming hybrid architecture vector engine, i.e. a VLIW architecture) and 20 SIPP accelerators (streaming image processing pipeline), which support a range of image processing functions. The SHAVE cores process data in 128-bit width.
Accelerates neural networks
To use the Neural Compute Stick, an x86 computer with a 64-bit processor and Ubuntu Linux 16.04 is required. The software development kit supplied with the stick is then installed on this development system. Another ingredient is the "Caffe" framework, which contains numerous algorithms for classifying and analyzing image data on the basis of neural networks. The Movidius SDK compiler translates the neural network generated by Caffe into machine instructions for the Myriad 2 processor.
The software must now be trained in the further development process. To do this, it is fed with real-time data from the real world and the results are compared with the calculations. Once trained, the stick can act as a discrete neural network accelerator on a prototype and expand an embedded system and local intelligence.













