Google Introduces its 8th-Gen AI Chip, the TPU 8T

Hinzugefügt: 2026-05-26

[00:00:00]And we've been investing for today and for the future.

[00:00:03]In 2022, we were spending $31 billion annually in CapEx.

[00:00:08]This year, we expect that number to be about six times that.

[00:00:12]Approximately 180 to 190 billion dollars.

[00:00:16]A key part of this investment is our custom silicon.

[00:00:20]A decade ago, we announced our very first commercial tensor processing unit or TPU on this I/O stage.

[00:00:27]Since then, we have transformed how the industry builds for AI.

[00:00:31]We recently announced our eighth generation of TPUs at Cloud Next.

[00:00:36]For the first time, we have taken a dual chip approach with specialized architectures for training and inference, TPU 8T and 8I.

[00:00:44]While they may look similar, they're actually pretty different.

[00:00:48]8T is optimized for large-scale pre-training and it's nearly three times the raw computing power of our previous generation.

[00:00:56]We have taken a fundamentally different approach with our training infrastructure.

[00:01:00]With JAX and Pathways, our training is no longer constrained by the limits of a single massive data center.

[00:01:07]Instead, we can now seamlessly distribute training across multiple sites, scaling across more than 1 million TPUs globally.

[00:01:16]This gives us the ability to create the largest training cluster in the world.

[00:01:20]Give you a live sense of what the speed feels like, here's a prompt on an upcoming flash model if it were running on 8I.

[00:01:29]I'll ask it to create a Chrome Dino game, push submit. The response is generated in real time.

[00:01:36]As you watch, take a look at the tokens per second in the top right corner. It almost took longer to write out the request and and the game is pretty fun, too.

[00:01:48]>> [applause] >> In addition to speed, we're also thinking about scaling sustainably.

[00:01:57]Both chips are more energy efficient, delivering up to two times better performance per watt. I bet Timmy TPU will be ready to teraflop right into bed after I/O.