Google TPU 8t / TPU 8i
Google's eighth TPU generation, split for the first time into a training chip and an inference chip.
You're looking at a chip record: what this processor is, its key specs, who designs, manufactures and supplies its parts, and where it gets used.
Announced at Google Cloud Next on April 22, 2026, the eighth TPU generation comes as two chips. TPU 8t targets training: 12.6 PFLOPS of FP4 and 216 GB of HBM per chip, in 9,600-chip superpods with about 2 PB of shared memory. TPU 8i targets serving and reinforcement learning: 10.1 PFLOPS of FP4, 288 GB of HBM and 384 MB of on-chip SRAM, linked by a new Boardfly topology. Google claims up to 2x performance per watt over Ironwood. In November 2026 it added TPU 8tx (confidential computing) and TPU 8i Max (two 8i dies per package); a general availability date is not confirmed.
Specifications & key facts
Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that: we never invent values.
Overview
Compute
Memory
Physical
Power
Manufacturing
How this connects to the rest of the stack
This record's links at a glance: click a group to jump to it.
Designed by
Cooling
Found inside
Uses
Available from
Successor to
Competes with
Related
Sources
Google Cloud: TPU 8t and TPU 8i technical deep dive (Apr 2026) ↗ · Google: Our eighth generation TPUs (Apr 2026) ↗ · Google Cloud: AI infrastructure at Next '26 (Apr 2026) ↗ · Google: Introducing TPU 8tx and TPU 8i Max (Nov 2026) ↗
Record reviewed as of 2026-11-16; see Methodology.