A supercomputer for neural computation

The requirement to train large neural networks quickly has prompted the design of a new massively parallel supercomputer using custom VLSI. This design features 128 processing nodes, communicating over a mesh network connected directly to the processor chip. Studies show peak performance in the range of 160 billion arithmetic operations per second. This paper presents the case for custom hardware that combines neural network-specific features with a general programmable machine architecture, and briefly describes the design in progress.