Chips · accelerators · Google
Google TPU v6e · Trillium
918 teraflops of dense BF16 compute a chip; Google publishes no power figure
Shipped 11 Dec 2024; shipped by the first date promised. Epoch AI names it at 14 AI data centers and no GPU cluster. Google across the site.
Embed
The frame credits Google, Epoch AI, AI data centers and Epoch AI, Data on GPU clusters and links back to its page. Preview it.
Cite
Each figure in Google’s own words
one chip, dense compute; memory is high-bandwidth memory- BF16, dense
- 918 TFLOPS
Peak compute per chip (bf16) 918 TFLOPs
TPU v6e - FP8, dense
- 918 TFLOPS
Peak compute per chip (FP8) (TFLOPs) 459 918 4614
TPU7x (Ironwood) (read from a table with a column for each product, whose head reads “Specification v5p v6e (Trillium) TPU7x (Ironwood)”) - Memory
- 32 GB
HBM capacity per chip 32 GB
TPU v6e - Memory bandwidth
- 1.64 TB/s
HBM bandwidth per chip 1638 GBps
TPU v6e (1638 GB/s on the page) - List price
- $2.70 a chip-hour
Trillium us-east1 South Carolina $2.70
Cloud TPU pricing (on-demand, us-east1 (first US region listed), per chip-hour); 340 TFLOP-hours of dense 16-bit compute a dollar
The v6e page gives Int8 (1836 TOPs) but no FP8; the FP8 figure (918, the second column) comes from the comparison table on the TPU7x page and equals the BF16 figure. Trillium went to preview on 2024-10-30.
1 dated promise and the shipment
each date links to the page it is quoted from- 15 May 2024 Promised Trillium TPUs available in 2024.
The sixth-generation Trillium TPUs are a culmination of over a decade of research and innovation and will be available later this year.
Announcing Trillium, the sixth generation of Google Cloud TPU - 11 Dec 2024 Shipped: Trillium generally available for Google Cloud customers.
Earlier this year, we announced Trillium, our sixth-generation and most performant TPU to date. Today, it is generally available for Google Cloud customers.
Announcing the general availability of Trillium, our sixth-generation TPU
Named at 14 AI data centers
H100-equivalents and power are the whole site’s today, with every chip it holds| Site | Owner | Country | H100e | MW | Chips |
|---|---|---|---|---|---|
| Google Columbus | United States | 331,986 | 364 | TPU v5e, TPU v5p, TPU v6e, TPU v7 | |
| Google New Albany | United States | 323,901 | 400 | TPU v5e, TPU v5p, TPU v6e | |
| Google Mesa | United States | 289,540 | 237 | TPU v6e, TPU v5p | |
| Google Council Bluffs (East) | United States | 255,685 | 284 | TPU v5e, TPU v6e, TPU v5p, TPU v7 | |
| Google Bristow | United States | 209,861 | 286 | TPU v5e, TPU v5p, TPU v6e, TPU v7 | |
| Google Omaha | United States | 198,585 | 284 | TPU v5e, TPU v5p, TPU v6e, TPU v7 | |
| Google Lincoln | United States | 197,575 | 170 | TPU v6e, TPU v7 | |
| Google The Dalles | United States | 175,341 | 185 | TPU v6e, TPU v5p, TPU v5e, TPU v7 | |
| Google Midlothian | United States | 143,581 | 123 | TPU v7, TPU v6e | |
| Google Storey County | United States | 133,906 | 214 | TPU v5e, TPU v4, TPU v6e, TPU v5p | |
| Google Lancaster | United States | 130,369 | 164 | TPU v5e, TPU v6e | |
| Google Waltham Cross | United Kingdom | 123,800 | 124 | TPU v6e | |
| Google Arcola | United States | 108,135 | 93 | TPU v6e, TPU v7 | |
| Google Red Oak | United States | 108,135 | 93 | TPU v6e |
In no GPU cluster of Epoch AI’s list
Epoch AI’s list of AI supercomputers, which stops in July 2025Epoch AI’s names for this chip: TPU v6e.
Other Google chips
- TPU v4 275 teraflops of dense BF16 compute a chip; Google publishes no power figure
- TPU v5e 197 teraflops of dense BF16 compute a chip; Google publishes no power figure
- TPU v5p 459 teraflops of dense BF16 compute a chip; Google publishes no power figure
- TPU7x (Ironwood) 2,307 teraflops of dense BF16 compute a chip; Google publishes no power figure
- Google’s own pages, each linked beside the figure or date it gives. Each page was read and each quote checked on 23 Oct 2026.
- Where it runs: Epoch AI, AI data centers and Epoch AI, GPU clusters, CC BY 4.0.