Infrastructure atlas
Google/Custom ASIC

Cloud TPU v6e / Trillium

Google’s sixth-generation TPU for transformer, text-to-image, and CNN training, fine-tuning, and serving.

TPUTPUtrainingservingJAX
Cloud TPU v6e / Trillium hardware

01 / Memory

32 GB HBM

Per chip

02 / Bandwidth

1,638 GB/s

HBM bandwidth

03 / Pod

256 chips

2D torus

Precision & peak compute

PrecisionPublished peakBasis
BF16918 TFLOPSPer chip
INT81,836 TOPSPer chip · AI TOPS

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

Google Cloud TPU v6e documentation

Technical specification

Chip & memory
Peak BF16
918 TFLOPS per chip
Peak INT8
1,836 TOPS per chip
HBM
32 GB · 1,638 GB/s
TensorCore
1 per chip · 2 MXUs
SparseCore
Included
Pod & interconnect
Pod size
256 chips
ICI
800 GB/s bidirectional per chip
Topology
2D torus
Host
8 chips · 1,536 GiB DRAM
Software
JAX, PyTorch/XLA, TensorFlow