SURVEY of ML accelerators

Relevant AI Chip Startups

Survey of Machine Learning Accelerators Albert Reuther

• The Syntiant NDP101 chip hSyntianti [51], [52] is the neural decision processor inside the Amazon Alexa products for attention phrase recognition. It is a processor-inmemory design, which performs inference weight computations in int4 and activation computation in int8 and
  consumes less than 200 µW.

• The Mythic Intelligent Processing Unit accelerator hMythici [53], [54] combines a RISC-V control processor, router, and flash memory that uses variable resistors (i.e., analog circuitry) to compute matrix multiplies. The accelerators are aiming for embedded, vision, and data center applications.

• The Perceive Ergo processor hPerceivei [55] is meant for battery-powered applications that require high performance. Ergo can run multiple concurrent networks
with over 100 million weights in applications such as simultaneous video and audio processing

AI 100: The Artificial Intelligence Startups Redefining Industries

Graphcore, Syntiant, Lightmatter, Xanadu, Neron