← All hardware
Google's 7th-gen TPU built from the ground up for the age of inference.
Pros
- Massive 9,216-chip single-pod scale
- Excellent perf/watt and price/performance via cloud
- Deep integration with Google Cloud and JAX
Cons
- Cloud-only; cannot be purchased on-prem
- Locked to Google Cloud ecosystem
- Smaller third-party software ecosystem than CUDA
✓ Where it shines / best for
- Large-scale training and inference on Google Cloud
- JAX/XLA-native research and production workloads
- Reasoning-model serving at hyperscale (e.g. Anthropic)
✕ Not the best fit for
- On-premises or multi-cloud-portable deployments
- CUDA-dependent codebases
- Small experiments better suited to single GPUs
Features
- ✓ 192 GB HBM per chip, ~7.37 TB/s bandwidth
- ✓ Pods scale to 9,216 chips via Inter-Chip Interconnect
- ✓ ~4.6 PFLOPS FP8 peak per chip
- ✓ 2x perf/watt vs TPU v6e Trillium
- ✓ Optical circuit switch (OCS) pod fabric
- ✓ Native JAX, PyTorch/XLA and TensorFlow support
- ✓ Liquid-cooled, cloud-only consumption model
- ✓ Optimized for reasoning and inference workloads
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| On-demand | Varies | per chip-hour | Consumption-based Google Cloud pricing; Google had not published a public flat per-chip-hour rate for tpu7x (Ironwood) as of this check. See Cloud TPU pricing page; access also via Dynamic Workload Scheduler. |
| Committed use / reservations | Discounted | 1-3 year | Lower rates with committed-use discounts and reservations |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| memory | 192 GB HBM per chip, ~7.37 TB/s |
| cooling | liquid-cooled, cloud-only |
| pod_scale | 9,216 chips per pod |
| architecture | Ironwood (TPU v7 / tpu7x), Google custom ASIC |
| perf_per_watt | 2x vs TPU v6e Trillium |
| fp8_performance | ~4.6 PFLOPS peak per chip |
Sponsored
A full review is being generated for this product and will appear here shortly.