Every AI chip on one chart: from 2 watts to a data centre

What is this chart?

MIT Lincoln Laboratory has kept a list of AI chips (called accelerators) since 2018. It plots each one by two numbers: how fast it can go at most, and how much power it needs at most.

MIT News wrote about the survey on 6 October 2026. The latest paper was posted on arXiv on 23 October 2025. The first paper covered 57 accelerators. The latest covers more than 120.

What does it show?

The authors sort the chips into five groups by where they are used: very low power (such as wake-word detection), embedded (cameras, small robots), autonomous (driver assistance), and two data-centre groups.

Both are quoted at int8 (numbers stored in 8 bits, which is enough for most AI that is already trained). These are the makers’ peak figures, not tests. They show the trade: more speed costs far more power.

Why should you care?

A small camera or robot has a small power budget. A data-centre chip has a far larger one. An AI idea has to be squeezed to fit the first.

The paper also says int8 is the usual choice for running trained models. Smaller numbers save energy.

What’s the catch?

• The chart uses public numbers only. Some companies do not publish theirs.
• Peak speed is not real speed. A program may reach only part of it.
• The survey covers chips for running trained models, mainly. Training is a different story.

Try it

Download the 2025 data file from the survey’s GitHub page. Plot the Power column against the PeakPerformance column on log scales. Pick out a chip you have heard of and see which group it falls in. If you share your chart, cite the survey’s papers, as its authors ask.

Sources

• Supercomputing researchers document evolution of AI hardware, MIT News, 6 October 2026
• Lincoln AI Computing Survey (LAICS) and Trends, Reuther and others, arXiv, 23 October 2025
• ai-accelerators: survey data and papers, GitHub