Published specifications
| Measure | Figure | Basis |
|---|
| Peak compute per chip | 123 TFLOPS BF16 | Current Google Cloud documentation[1][2] |
|---|
| HBM2 capacity | 32 GiB per chip | Nominal capacity[1][2] |
|---|
| HBM bandwidth | 900 GB/s per chip | Maker specification[1][2] |
|---|
| Pod size |
|---|
Research to commercialization
- Announced
- Cloud beta
- Cloud GA
- Pod GA
May 8, 2018Demonstrated
TPU 3.0 revealed at Google I/O[5]
Oct 10, 2018Demonstrated
Cloud TPU v3 beta[1][2]
Jan 31, 2019Demonstrated
Cloud TPU v3 general availability
Contemporary alternatives — different platforms and workloads
| Product | Published context |
|---|
| Tesla V100 | Volta training and inference GPU, announced May 10, 2017; maker launch states 120 deep-learning teraflops.[3][4] |
|---|
| Tesla P100 | Pascal data-center accelerator announced April 2016; a contemporary general-purpose GPU alternative.[3][4] |
|---|
Product context, not a benchmark ranking. GPU, TPU and cloud-instance peak rates and memory units may differ.
