Chips · accelerators · Google
Google TPU v4 · TPU
275 teraflops of dense BF16 compute a chip; Google publishes no power figure
Shipped 12 Oct 2022; shipped after the first date promised, up to 10 months. Epoch AI names it at 2 AI data centers and 4 GPU clusters. 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
- 275 TFLOPS
Peak compute per chip 275 teraflops (bf16 or int8)
TPU v4 - INT8, dense
- 275 TOPS
Peak compute per chip 275 teraflops (bf16 or int8)
TPU v4 - Memory
- 32 GiB
HBM2 capacity and bandwidth 32 GiB, 1200 GBps
TPU v4 - Memory bandwidth
- 1.2 TB/s
HBM2 capacity and bandwidth 32 GiB, 1200 GBps
TPU v4 (1200 GB/s on the page) - List price
- $3.22 a chip-hour
TPU v4 pod us-central2 Oklahoma $3.22
Cloud TPU pricing (on-demand, us-central2 (Oklahoma), TPU v4 pod, per chip-hour); 85 TFLOP-hours of dense 16-bit compute a dollar
Google does not publish a TDP for TPU v4; the v4 page gives only measured power (90/170/192 W min/mean/max), so power is left out. TPU v4 Pods went to preview on 2022-05-12 and to general availability on 2022-10-12, after the 2021 window.
1 dated promise and the shipment
each date links to the page it is quoted from- 18 May 2021 Promised TPU v4 pods available to Cloud customers in 2021.
They'll be available to our Cloud customers later this year.
Google I/O 2021: Being helpful in moments that matter - 12 Oct 2022 Shipped: Cloud TPU v4 Pods generally available.
Cloud TPU v4 Pods - Now in General Availability (GA)
Google Cloud infrastructure enhancements tailored for your workloads
Named at 2 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 Storey County | United States | 133,906 | 214 | TPU v5e, TPU v4, TPU v6e, TPU v5p | |
| Google Papillion | United States | 100,556 | 284 | TPU v4, TPU v5e, TPU v5p |
71,680 chips in 4 GPU clusters
Epoch AI’s list of AI supercomputers, which stops in July 2025| Cluster | Owner | Country | Chips | First operational |
|---|---|---|---|---|
| Google Oklahoma TPU v4 Pods | United States of America | 32,768 | 2022-05-12 | |
| Gemini 1.0 Ultra training cluster A | United States of America | 28,672 | 2023-08-20 | |
| Paper on PaLM | United States of America | 6,144 | 2022-04-05 | |
| Google TPU v4 Pod | United States of America | 4,096 | 2021-05-19 |
Epoch AI’s names for this chip: TPU v4; Google TPU v4.
Other Google chips
- 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
- TPU v6e 918 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.