Google TPU7x (Ironwood)
Google TPU7x is the first release of Ironwood, Google’s seventh-generation TPU, available on Google Cloud.
- Vendor
- Architecture
- Ironwood (7th-gen TPU), dual chiplet
- Memory
- 192 GB HBM
- Memory bandwidth (TB/s)
- 7.37 TB/s
- FP8 (TFLOPS)
- 4,614 TFLOPS
- Form factor
- Cloud TPU (4-chip VMs, pods up to 9,216 chips)
- Status
- Shipping
TPU7x is available on Google Cloud through Compute Engine and GKE. Each chip has two chiplets (two TensorCores, four SparseCores) with 192 GB of HBM at approximately 7.37 TB/s, and pods scale to 9,216 chips in a 3D torus. Every TPU7x slice uses full-host VMs of four chips. Google lists peak FP8 compute per chip as 4,614 TFLOPS and BF16 as 2,307 TFLOPS.
- Memory (GB)
- 192 GB
- Precision note (sparse or dense)
- Google lists peak FP8 compute per chip as 4,614 TFLOPS (BF16 2,307) with no sparse/dense label. Bandwidth stated as approximately 7.37 TB/s (7,380 GBps in Google's table).
- Vendor source
- cloud.google.com
- Checked
Questions
What is Google Ironwood?
Ironwood is Google’s seventh-generation TPU, and TPU7x is its first release, available through Compute Engine and GKE.
How much memory does TPU7x have?
Each TPU7x chip has 192 GB of HBM, with bandwidth of approximately 7.37 TB/s.
What is TPU7x’s FP8 performance?
Google lists peak FP8 compute per chip as 4,614 TFLOPS (BF16 2,307), with no sparse or dense label.
How many chips are in an Ironwood pod?
TPU7x pods scale to 9,216 chips in a 3D torus, and every slice uses full-host VMs of four chips.
Source
Every figure is taken from the vendor’s own datasheet, product page, technical documentation or announcement, linked on each chip with the date it was last checked. How we pick and check sources: editorial standards.