MIT News breed · Science

AI Hardware Evolution Documented

Researchers are documenting the rapid evolution of AI hardware accelerators through the Lincoln AI Computing Survey (LAICS), aiming to clarify the advancing field of specialized AI technology.

The LAICS initiative, started in 2018, tracks AI accelerators crucial for technological advantage. It began due to government interest in these systems.

Accelerators include CPUs, GPUs, ASICs, FPGAs, and dataflow types, each with unique capabilities and varying performance/efficiency.

The survey compares accelerators by peak performance and power, using public data. The number of tracked accelerators grew from 57 to over 120.

New startups and innovative accelerators continue to emerge annually, indicating a dynamic market.

LAICS papers also explore performance drivers like smaller transistors and lower numerical precision, and analyze architectural choices.

The team collaborates to understand accelerator use in research and development, planning to continue the survey due to rapid innovation.

The survey aids sponsors and government colleagues in making informed decisions about AI technologies and acquisitions.

AI-samenvatting op basis van de bron.