Our eighth generation TPUs: two chips for the agentic era
Article
TL;DR: Google splits TPU 8 into dedicated inference (8i) and training (8t) chips, delivering 2x perf/watt over gen 7.
Key Takeaways
- TPU 8t superpod: 9,600 chips, 2PB shared HBM, 121 ExaFlops of compute
- Vertical integration lets Google optimize cost at datacenter scale — Nvidia can’t match that
- At $15/GB for HBM4, the 331TB per pod is ~$5M in memory alone
Discussion
Top comments:
- [pmb]: Google’s whole-datacenter context makes it structurally more cost-efficient than chip vendors
- [TheMrZZ]: 9,600 chips and 2PB shared HBM sounds like a real competitive advantage
- [Keyframe]: Google’s been growing quietly while competitors dominated the news cycle
- [paulmist]: At $15/GB HBM4, the 331TB per pod is $5M in memory alone