DEPLOY

AI chip · Google

Google TPU v4

Fourth-generation Tensor Processing Unit; optical-switch-networked pods, the foundation for PaLM / Gemini training through 2024. Fabricated by TSMC.

Deployed in 2 named data centers.

Market position

TPU v4 ran the PaLM and Gemini-Nano-generation training. Its 4,096-chip pod with optical circuit switching was Google's answer to InfiniBand fat-tree. Still deployed at The Dalles and Mayes County; superseded by v5p for new training workloads.

Efficiency and power

How much work you get per watt and per dollar, and how much power a full rack draws.

Perf per watt
1.43 BF16 TFLOPS/W
275 TFLOPS ÷ 192 W = 1.43

What fits in 32 GB

Which open-source LLMs run on one Google TPU v4, by precision. Weights only: add roughly 15% headroom for real serving. If a model does not fit at FP16, try INT8 or INT4 (smaller quality trade-off than most people expect).

ModelParamsFP16INT8INT4
Llama 3.1 8B
dense
8 B16 GB ✓8 GB ✓4 GB ✓
Llama 3.1 70B
dense
70 B140 GB ✗70 GB ✗35 GB ✗
Llama 3.1 405B
dense
405 B810 GB ✗405 GB ✗203 GB ✗
Llama 3.3 70B
dense
70 B140 GB ✗70 GB ✗35 GB ✗
DeepSeek V3
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
DeepSeek R1
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
Qwen 2.5 7B
dense
7 B14 GB ✓7 GB ✓4 GB ✓
Qwen 2.5 72B
dense
72 B144 GB ✗72 GB ✗36 GB ✗
Mixtral 8x7B
MoE (46.7B total, 12.9B active per token)
46.7 B93 GB ✗47 GB ✗23 GB ✓
Mixtral 8x22B
MoE (141B total, 39B active per token)
141 B282 GB ✗141 GB ✗71 GB ✗
Gemma 2 27B
dense
27 B54 GB ✗27 GB ✓14 GB ✓
Command R+
dense
104 B208 GB ✗104 GB ✗52 GB ✗
Kimi K2
MoE (1T total, 32B active per token)
1000 B2000 GB ✗1000 GB ✗500 GB ✗

Math: FP16 = params × 2 bytes; INT8 = params × 1 byte; INT4 = params × 0.5 bytes. MoE models sum every expert (full weights on disk), not the per-token active subset.

Common questions

How much does Google TPU v4 cost?

Google TPU v4 doesn't have a public launch list price. Vendors like this one usually price through direct sales rather than a public sheet. See list price per TFLOP for chips that do publish.

How much memory does Google TPU v4 have?

32 GB of HBM2, running at 1,200 GB/s. Straight from the vendor datasheet. See chips with the most memory for context.

How much power does one Google TPU v4 draw?

192 W at the chip. A full server draws more once you add CPU, memory, networking and cooling: see the rack power number above. Compare to other chips on perf-per-watt.

Which data centers use Google TPU v4?

2 named data centers run them, including Google Pryor (MidAmerica) and Google The Dalles. See who has the most Google TPU v4 for the ranked list.

Which open-source LLMs fit on one Google TPU v4?

Of 13 open-source LLMs we track, 2 fit at FP16, 3 at INT8, and 4 at INT4 (for example Llama 3.1 8B, Qwen 2.5 7B and Mixtral 8x7B). Full table above with each model's memory need.

See every answer we publish →

Key facts

Class
Compute SoC (AI accelerator)
Designer
Google
Safety-critical?
No (data-center inference / training)
Record as of
2026-09-27
Most recent source
2024-01-01 (across 2 sources on this page)
Specifications (9 fields, click to expand)

Straight from the vendor datasheet. Dense throughput shown first; sparse (2:4) numbers in parentheses where the vendor publishes them. Full datasheet linked below.

Process node
TSMC N7
TDP
192 W
Memory
32 GB HBM2
Memory bandwidth
1,200 GB/s
BF16 (dense)
275 TFLOPS
INT8 (dense)
275 TOPS
Form factor
OAM (per-chip in v4 pod)
Announced
2021-05-18
Released
2022-05-01

Source: vendor datasheet

Generation

Foundry & process

Fabbed at
TSMC (all chips TSMC makes →)
Process node
TSMC N7

Compare with

See every chip comparison →

Data centers running Google TPU v4

Who has the most? →

Named data-center campuses with Google TPU v4 on site. Counts shown where the operator has published them; other rows are described in general terms.

  1. partially energizedreported

    us-central2 TPU v4 pods (Google publishes region availability)

    Google Cloud

  2. partially energizedreported

    TPU v4 pods (major Google Cloud TPU region)

    Google Cloud

Sources

Adoption rows appear as we confirm each chip-in-robot pairing from a public source. See every chip we track for the full catalog or the Google page.