Lincoln Laboratory continues survey tracking AI accelerators

Summary

The MIT Lincoln Laboratory Supercomputer Center has been conducting the Lincoln AI Computing Study (LAICS) since 2018. The center has published six papers cataloging and comparing the peak performance and power consumption of over 120 commercial AI accelerators to guide purchasing decisions for the laboratory and its government sponsors.

Why it matters

As the AI accelerator market fragments across CPUs, GPUs, ASICs, FPGAs, and dataflow chips, regular, structured comparisons help defense and research buyers look beyond vendor marketing. For a laboratory serving national security missions, a continuously updated, publicly documented benchmark serves as a practical insurance policy against reliance on any single vendor's roadmap.

Highlights

• Six papers have been published since 2018. The latest version covers over 120 accelerators, compared to the 57 in the initial release.
• Core metrics remain the comparison of peak performance versus peak power consumption. This comparison is segmented by chip, card, or system form factor.
• Reuther reports that 5 to 10 new accelerator initiatives emerge each year, contributing to the sector's dynamic and ever-changing nature.

Critical take

The study relies solely on publicly available data. This means the most sensitive performance data, which defense-critical vendors keep confidential, is systematically overlooked. Furthermore, a two-metric comparison of peak performance and peak power consumption provides little insight into real-world data transfer rates, latency, or software stack maturity—factors that actually determine how well an accelerator performs for a specific workload.