Lincoln laboratory publishes survey of AI accelerators

LAICS uses peak performance and power as its main metrics, then sorts devices by form factor—chip, card or system. The survey also explores how architectural choices, such as adding cores or parallel performance, affect performance. In the latest paper, the team noted that smaller, denser transistor designs and lower numerical precision drive performance gains. Reuther reports that each year five to ten new startups launch accelerators, and six have announced their first products in recent months.

The survey informs LLSC’s procurement decisions, helping the laboratory choose GPUs for future systems and advising government sponsors on acquisition choices. Reuther says the data help sponsors and mission programs understand the AI accelerator landscape and make better research and acquisition decisions. The full set of papers and datasets is available online.