Google expanded its eighth-generation TPU lineup by introducing the TPU 8tx and TPU 8i Max. The TPU 8tx pairs TPU 8t silicon with 5th-gen AMD Epyc processors for confidential AI, while the TPU 8i Max combines two 8i chips to handle massive memory workloads.
The TPU 8tx uses Boardfly topology to double performance over the 8t, securing data and weights during processing. The TPU 8i Max doubles compute power and HBM bandwidth to 17,202 GB/s, featuring 768 MB of on-chip SRAM and 576 GB of HBM for large contexts.
New software tools like TorchTPU and the XLA Serving Engine optimize these systems for PyTorch and multi-hardware serving. With Google Cloud Pathways and dynamic slicing now generally available on GKE, firms can more efficiently manage large TPU clusters.