Lincoln Lab Survey Tracks 120-Plus AI Accelerators as 6 Papers Chart 8 Years of Hardware Shifts
Updated
Updated · MIT News · Oct 6
Lincoln Lab Survey Tracks 120-Plus AI Accelerators as 6 Papers Chart 8 Years of Hardware Shifts
1 articles · Updated · MIT News · Oct 6
Summary
Lincoln Laboratory’s latest LAICS paper compares more than 120 commercial AI accelerators, up from 57 in the first survey, extending a series that has now reached six papers since 2018.
The survey was launched after government sponsors asked about a fast-growing accelerator market, and it benchmarks chips, cards and systems mainly by peak performance and power using public data.
The newest paper goes beyond cataloging devices to test architectural trade-offs, examining how choices such as more cores per processor and greater parallelism could change system performance.
Albert Reuther said the field is still expanding, with another five to 10 startups appearing each year and six new startups announcing first accelerators in recent months.
Lincoln Lab says the survey helps sponsors and government users make research and acquisition decisions, while also guiding its own future GPU purchases for supercomputing systems.