Trillium, the sixth generation of Google Cloud TPU
cloud.google.com
cloud.google.com
But I dug up this:
TPUv5e has 197 (bf16) TOPS (Tera Operations Per Second)
TPU V6 is 4.7x that, so...
TPU V6 is maybe 925.9 (bf16) TOPS
And I think I'm comparing apples to apples when I see this;
H200 BFLOAT16 Tensor Core 1,979 TFLOPS
So from my napkin calculations, Nvidia is 2x the raw flops... But I didn't compare memory, bandwidth, or power efficiency... Or cost.