What Machine Learning Needs from Silicon

Wednesday, July 10
2:05pm to 2:35pm

This talk will cover the unique requirements of machine learning workloads, and how they translate into demands on the underlying hardware. This will include the high intensity of arithmetic operations compared to memory, the loose precision requirements, how much storage is needed for model parameters. These workloads have unusual properties like a complete lack of data-driven branches, and knowledge of exact memory access patterns for many millions of cycles into the future. We will also discuss how occasional general-purpose logic operations in the middle of a sequence of large-scale numerical calculations can hurt performance if they require a round-trip between a neural accelerator and an application processor.

Share page with AddThis