Cloud TPU v5p
TPU v5p pods of 8,960 chips, 2x the FLOPS and 3x the HBM of TPU v4, announced with Gemini.
Google's most powerful TPU at the time, with 2.8x faster LLM training than v4 and 4x more FLOPs per pod. It was used to train Gemini and came packaged with the 'AI Hypercomputer' software stack.
- Date
- Wednesday, 6 December 2023
- Lab
- Kind
- infra
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| LLM training speed vs TPU v4 | 2.8x pod of 8,960 chips | company |
Gemini launch post (2023-12-06) announced v5p; the Cloud blog entry is dated 2023-12-07.
Sources
- cloud.google.com/blog/products/ai-machine-learning/introducing-cloud-tpu-v5p-and-ai-hyperc
- blog.google/technology/ai/google-gemini-ai/
This record was checked against its sources on 6 October 2026. How we check