MIT survey maps the changing market for AI accelerators
MIT Lincoln Laboratory expands its AI hardware survey
MIT Lincoln Laboratory’s latest Lincoln AI Computing Survey (LAICS) examines more than 120 commercial AI accelerators, compared with 57 in the first paper in the series. The work is led by Albert Reuther at the Lincoln Laboratory Supercomputing Center (LLSC), whose team has tracked the market since 2018.
The sixth LAICS paper compares accelerators using peak performance and peak power, then groups them by whether they are delivered as a chip, card or complete system. The survey draws its data from public material, including technical press coverage, company announcements and industry presentations.
Reuther said the project began as research papers and commercial announcements rapidly multiplied, prompting questions from government sponsors. He runs daily news and citation searches to identify new material, an effort made necessary in part because some vendors do not publish performance and power data.
Different architectures serve different workloads
The survey covers central processing units, graphics processing units, application-specific integrated circuits, field-programmable gate arrays and dataflow accelerators. CPUs support general-purpose computing, while ASICs are built for highly specific tasks. GPUs, FPGAs and dataflow designs offer varying degrees of flexibility for multiple workloads.
These systems are often associated with neural networks, deep learning and machine learning, but the laboratory notes that they can also support parallel work such as molecular-function modelling and fluid-dynamics simulation. Performance and efficiency vary with the design of each accelerator and with the work it is asked to perform.
Each LAICS publication adds a different analytical angle. A 2022 paper studied the sources of performance gains and identified smaller, denser transistor designs and lower numerical precision as contributors. The newest paper examines architectural options, including the effect of adding components such as more processor cores or parallel performance.
Procurement decisions need more than a headline metric
LAICS is intended to help Lincoln Laboratory sponsors and government colleagues understand the expanding accelerator landscape and make research and acquisition decisions. Reuther said that six new startups had announced their first AI accelerators in only the preceding few months, underscoring the pace of new entries.
The findings also inform upcoming LLSC system purchases, including the GPUs the center may consider. For organisations assessing AI infrastructure, the practical implication is to compare publicly available peak performance and power figures alongside architecture, form factor and the requirements of the intended workload before making an acquisition choice.