Google split 8th-gen TPU into separate training (8t) and inference (8i) chips; first time separating these tasks
Awaiting review — this entry has not been approved yet.
Details
occurred at
2026-04-22
summary
First time Google splits training/inference into separate TPUs; inference chip uses SRAM like Groq/Cerebras; 384MB SRAM per chip; training chip 2.8x Ironwood; DA Davidson valued TPU+DeepMind at ~$900B