Intel Announces Myriad X, a Dedicated Chip that Can Handle a Deep Neural Network Independently with a Minimal Size

Intelneural networkSpecial processor dedicated to processing of "Myriad XWe announced. Conventional chip ratio that can be 10 times the neural network processing Myriad X is, take advantage of to, such as robots and autonomous flying drones is expected.

Intel Unveils Neural Compute Engine in Movidius Myriad X VPU to Unleash AI at the Edge | Intel Newsroom
https://newsroom.intel.com/news/intel-unveils-neural-compute-engine-movidius-myriad-x-vpu-unleash-ai-edge/

Intel Unleashes Myriad X Vision Processing Unit With Neural Compute Engine And AI Capabilities
https://www.forbes.com/sites/marcochiappetta/2017/08/28/intel-unleashes-myriad-x-vision-processing-unit-with-neural-compute-engine-and-ai-capabilities/

Myriad X is a processor named "vision processing unit (VPN)" designed to handle neural networks, not just video processing. It is a tiny tip of 8.7 mm × 8.5 mm size, and Movidius under the umbrella of Intel developed it.


"The use of deep learning is becoming a standard condition in computing environments," said Remi Ellusen, vice president of Movidius. "Enabling a terminal that can perform visual processing like human beings means the next leap of computing We are saying that SoC's Myriad X is required to deeply deepen the deep learning process which can be processed in the terminal rather than the cloud base.


The arrangement of CPU and memory in Myriad X can be confirmed with the following movie.

Myriad ™ X: Chip Anatomy - YouTube


Myriad X is the world's first SoC developed for deep learning processing. Myriad X's 16 SHAVE (Streaming Hybrid Architecture Vector Engine) processors consist of a programmable 128-bit VLIW vector unit. The inference performance of DNN is 1 TOPS, it can execute deep neural network with high speed and low power.

In addition, by adopting an extended video accelerator and an on-chip memory architecture with an internal bandwidth of 400 GB / s, it reduces the number of data transfers with the outside, minimizes latency and suppresses power consumption. Multiple images / images can be executed simultaneously in the pipeline, and up to 8 HD resolution RGB cameras are connected using 16 MIPI lanes, achieving signal processing throughput of 700 million pixels per second. The new hardware encoder supports 4 K resolution of 30 fps with H.265 / H.264 and 60 fps with M / JPEG. Myriad X also includes USB 3.1 and PIC-Express Gen 3 interface.


Myriad X said that two versions of "MA 2485" including 4 Gbit LPDDR 4 and "MA 2085" without internal memory will be prepared. Myriad X is expected to be used for drones, surveillance cameras, robots, VR / AR terminals, and so on.