Google TPU v4 Pod
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| BF16 | 1126.4 | 275.0 | — |
| INT8 | 1126.4 | 275.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- TPU v4 32GB
- Count
- 4096
- Memory per GPU
- 32 GB
- Bandwidth per GPU
- 1.2 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- —
- BF16 Efficiency
- —
- Aggregate Bandwidth
- 4915.2 TB/s
System Details
Interconnect
3D mesh ICI, 1.1 PB/s all-reduce bandwidth per pod, 24 TB/s bisection bandwidth
Notes
A full TPU v4 pod is 4,096 chips in a 3D mesh. Google publishes a single peak compute figure of 275 teraflops per chip covering both bf16 and int8 rather than separate rows, giving 1,126.4 PFLOPS across the pod; Google prints this rounded as "1.1 exaflops (bf16 or int8)". Also published per pod: 1.1 PB/s all-reduce bandwidth and 24 TB/s bisection bandwidth. Memory is 4,096 x 32 GiB HBM2 at 1,200 GBps per chip. Google measures per-chip power at 90 W minimum, 170 W mean and 192 W maximum, but publishes no pod-level power, so total system power is left blank here rather than extrapolated from silicon alone. Google publishes no sparse figures for any TPU, so all figures are dense.