Google TPU v4
TPU v4 Architecture
Active
Launched May 2021
Core Specifications
VendorGoogle
ArchitectureTPU v4
Form FactorMezzanine
VRAM—
Memory Bandwidth—
TDP300 W
Compute Performance
| Precision | TFLOPs |
|---|---|
| BF16 | 275 |
Performance Benchmarks
llm train
llm inference
| Configuration | Precision | Performance | Source |
|---|---|---|---|
| Gemini models, batch inference | BF16 | 6,500 tokens_per_second | View |
Pricing
Cloud Rental (OPEX)
| Provider | Instance Type | Price per Hour | Region | As of |
|---|---|---|---|---|
| Google Cloud | — | $3.00/hr | us-central | Dec 2025 |