# Google unveils two new Tensor Processing Units.

*semiconductor · news · 2026-04-22 · Interesting Engineering*

## Key points

- Google introduced two distinct AI chips, TPU 8t for training and TPU 8i for inference.
- TPU 8t achieves nearly triple the compute performance per pod over the previous Ironwood generation.
- TPU 8i offers 80% better performance-per-dollar than its predecessor and includes 288 GB high-bandwidth memory.
- Both chips provide up to two times better performance-per-watt compared to Ironwood and use fourth-generation liquid cooling.
- TPU 8t and 8i support major AI frameworks like JAX, PyTorch, SGLang, and vLLM for seamless deployment.

**Companies:** Google
**Countries:** United States

[Read the full story on Interesting Engineering](https://interestingengineering.com/ai-robotics/google-tpu-8t-8i-ai-chips-launch)

---

Canonical: https://newsio.io/n/c32df065-8075-43f9-bf9f-14d1924e6143/google-unveils-two-new-tensor-processing-units-tpus-are-in-house-accelerators-th
Summarized by Newsio from Interesting Engineering. https://newsio.io/how-it-works