Hardware

The inference half of the split, with up to 80% better performance per dollar than Ironwood and 3x the on-chip SRAM.

AWS’s first in-house machine-learning chip, for low-cost inference.

4.7x the peak compute per chip of TPU v5e, with double the HBM capacity and bandwidth.

Built for inference, with 6x Trillium’s HBM capacity and 4.5x its bandwidth.

Google split the line in two; the training chip claims up to 2.7x performance per dollar over Ironwood.

Cost-focused TPU with up to 2x training and 2.5x inference performance per dollar vs TPU v4.

Training TPU with more than 2x the FLOPS and 3x the HBM of TPU v4.

Moves to 7nm and triples in-package memory to 96 GB of HBM2E.

Moves to 5nm with 4x the BF16 AI compute and 1.5x the memory bandwidth of Gaudi 2.

About 2.2x the peak FLOPs of TPU v3 per chip.