Graphcore Intelligence Processing Unit

A new yew—IPU Graphcore continues to dribble out information about its hopefully soon-to-be-released ‘Intelligence Processing Unit’. Earlier this year we got past the It’s-going-to-be-awesome stage, and we’ve since learned it has enough memory to hold an entire neural network, including the network’s weights, on chip for at least the duration of an inference. They carefully avoid telling us how large a network they can accommodate in on-chip memory but they do tell us that it delivers 70 TBytes/sec of aggregate bandwidth. Memory and compute cores are connected by a many-to-many network fabric on chip that is accessible to the compiler.