A TPU isn't a faster GPU — it's a different chip

Added: 2026-05-19

[00:00:00]So, what is a TPU? Here's the one idea, and once you see it, the whole story makes sense. A GPU is thousands of tiny general-purpose cores, flexible, but every one of them burns time fetching instructions, decoding them, shuffling data around, overhead everywhere. A TPU throws all of that out. At its heart sits one enormous grid of multiply-add cells, a systolic array.

[00:00:26]The first TPU's grid was 256 by 256.

[00:00:30]That's 65,000 little multipliers. You load the weights into the grid once, then you pump the data through. It flows across, partial sums fall downward, the answers drop out the bottom, like a heartbeat. Almost every transistor is doing actual math, no fetch, no decode, no waste, one job done perfectly.