Researchers at MIT Lincoln Laboratory are keeping a running record of the hardware that powers artificial intelligence. The work was described in a report published on October 6, 2026.
The project is called the Lincoln AI Computing Survey, or LAICS, pronounced "lace." A team from the Lincoln Laboratory Supercomputing Center has run it since 2018.
The survey compares commercial AI accelerators. These are specialized systems built to speed up tasks such as neural networks, deep learning, and machine learning.
Why the Survey Started
Albert Reuther, a staff member at the center, leads the effort. He said the team saw a sharp rise in research and commercial accelerators about eight years ago.
Government sponsors of the laboratory began asking questions about the new hardware. Reuther said that was "motivation enough to start the survey."
The team also includes Michael Jones, Peter Michaleas, Jeremy Kepner, and Vijay Gadepally. They work with other groups at the laboratory to learn how accelerators support research missions.
AI accelerators are not used only for machine learning. They can also help model molecules and speed up fluid dynamics simulations, which take a lot of computing power.
How the Hardware Is Compared
Accelerators come in several forms. These include CPUs, GPUs, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and dataflow accelerators.
CPUs handle general computing, while ASICs perform only very specific tasks. GPUs, FPGAs, and dataflow accelerators are more flexible and can be set up for many workloads.
The team's main measures are peak performance and peak power. Accelerators are then sorted by whether they are a chip, a card, or a full system.
All data comes from public sources. Reuther said this can be hard because some companies keep their performance and power numbers private.
To stay current, Reuther runs daily news and citation searches. These track technical articles, company announcements, and industry presentations.
The survey has grown over six papers. The first studied 57 accelerators, and the latest covers more than 120.
Each paper also looks at a new topic. The 2022 paper found that performance gains came from smaller, denser transistors and the use of lower numerical precision.
The latest paper studied design choices, such as adding more cores per processor. It looked at how those changes would affect a system.
Reuther said the survey helps sponsors make better research and buying decisions. It also helps the center decide which GPUs to consider for its own future systems.
"It continues to surprise me how each year another five to 10 startups get funded and announced, and then release new AI accelerators," Reuther said.
Reuther plans to keep the survey going. He said six new startups have announced their first AI accelerators in just the past few months.
