Intel’s 1Tops neural processor for vision

Intel has introduced a neural computing engine for accelerating deep learning inferences at the edge.

It is included on an SoC called Myriad X, aimed at drones, robotics, smart cameras and virtual reality.

“The Myriad X architecture is capable of 1Tops maximum performance, based on peak floating-point computational throughput on deep neural network inferences,” said the firm.

Also on-board are:

16 programmable 128bit VLIW vector processors optimised for computer vision workloads, for running multiple imaging and vision application pipelines simultaneously.
16 MIPI lanes to connect up to 8 eight high resolution RGB cameras to Myriad X, supporting up to 700Mpixel/s image signal processing throughput.
“Over 20” said Intel, hardware vision accelerators to perform tasks such as optical flow and stereo depth without introducing additional compute overhead.
2.5Mbyte of homogenous on-chip memory, with up to 450Gbyte/s bandwidth.

According to the firm, 4Tops of processing is available all told.

“Myriad X is the newest generation in a lineage of Movidius vision processing units, which are purpose-built for embedded visual intelligence and inference,” said Intel. “Movidius VPUs achieve significant performance at low power with the merging of three architectural elements to provide sustained high-performance on deep learning and computer vision workloads: an array of programmable VLIW vector processors with an instruction set tuned to computer vision and deep learning workloads; a collection of hardware accelerators supporting image signal processing, computer vision, and deep learning inferences; and commonly accessible intelligent memory fabric that minimises data movement on chip.”

The chip is 8.1 x 8.8mm.