← All hardware
Google's widely available 6th-gen TPU for efficient training and inference.
Pros
- Generally available and widely proven
- Cost-efficient training/inference via cloud
- Tight Google Cloud and JAX integration
Cons
- Lower per-chip memory than v7 Ironwood
- Cloud-only, Google-ecosystem-locked
- Superseded at the high end by Ironwood
✓ Where it shines / best for
- Cost-efficient transformer training on GCP
- High-throughput inference serving
- Teams already on JAX/XLA
✕ Not the best fit for
- On-prem or portable multi-cloud needs
- Largest frontier models needing v7 scale
- CUDA-only software
Features
- ✓ 32 GB HBM per chip, ~1,600 GB/s bandwidth
- ✓ 256-chip pod footprint with ICI fabric
- ✓ 4x training and up to 3x inference vs prior gen
- ✓ Optimized for transformer training and serving
- ✓ Native JAX, PyTorch/XLA, TensorFlow support
- ✓ Generally available across Google Cloud regions
- ✓ Multislice training across pods
- ✓ Strong perf/watt efficiency
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| On-demand | ~$2.70 | per chip-hour | On-demand list price ~$2.70/chip-hour (varies by region); consumption-based |
| 1-year commitment | ~$1.89 | per chip-hour | Committed-use discount (1-year) |
| 3-year commitment / Spot | ~$1.22 (3yr); Spot ~$0.60-$1.30 | per chip-hour | 3-year committed-use is further discounted; Spot/preemptible ~$0.60-$1.30/chip-hour |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| memory | 32 GB HBM per chip, ~1,600 GB/s |
| cooling | cloud-only |
| pod_scale | 256 chips per pod |
| performance | 4x training / up to 3x inference vs prior gen |
| architecture | Trillium (TPU v6e), Google custom ASIC |
Sponsored
A full review is being generated for this product and will appear here shortly.