Google adds TPU 8tx and 8i Max, taking its eighth-generation TPU family to four chips
Google has expanded its eighth-generation TPU line with TPU 8tx, for confidential inference and small-scale training, and TPU 8i Max, a dual-chip package for long-context inference. It also open-sourced TorchTPU and TPU Sync. Availability dates for the new chips were not given.
Google announced two new chips on 9 November, bringing its eighth-generation TPU family to four. TPU 8t targets large-scale training and TPU 8i targets real-time reasoning, with the new pair filling gaps at the secure and long-context ends.
TPU 8tx uses the same silicon as TPU 8t but pairs it with 5th Gen AMD EPYC host CPUs for confidential workloads with hardware-level encryption. Google says it delivers up to twice as low a latency as TPU 8t, and that regulated customers asked for the encryption capability.
TPU 8i Max places two TPU 8i chips on one package, doubling compute, SRAM and HBM bandwidth to suit long-context inference, within the existing rack footprint.
On software, Google open-sourced TorchTPU, which runs native PyTorch and Hugging Face models on TPUs, and TPU Sync, which it says cuts KV-transfer latency and time to first token by more than 90 per cent. GKE dynamic slicing is now generally available. Google says nearly 100 customers use TPUs. It did not publish availability dates for the new chips.
The figures are Google's own claims and have not been independently tested.
Why it matters
Inference, not training, is increasingly where AI costs fall. Hardware aimed at confidential computing speaks directly to banks, health providers and governments that have held back sensitive workloads.
Indicops view
Indian government and regulated-sector buyers should note the confidential-compute direction and the PyTorch compatibility, which lowers switching cost away from GPU-only stacks. Ask for regional availability and pricing before committing roadmaps to hardware without published dates.
More in AI News
Back to the newsroomJensen Huang steps down as Nvidia CEO; CFO Colette Kress becomes interim chief
Nvidia's founder resigned as chief executive on 12 November after a board meeting, with CFO Colette Kress taking over on an interim basis while a search begins. The shares fell as much as 13 per cent, and quarterly results are due on 17 November.
Iran-linked hackers sabotage refrigeration controllers across Israel, the US and Europe
NPR reports that a group calling itself Cyber Isnaad Front, which Israeli incident responders attribute to Iranian military hackers, has sabotaged industrial refrigeration systems from a food plant in Israel to roughly 300 French supermarkets and, in early November, retailers across six more European countries. Internet-exposed Danfoss controllers were a common entry point.
HSBC warns India's data-centre build-out could add $105 billion to the import bill
HSBC Global Research estimates that India's planned jump from roughly 1.5 GW to 6.5 GW of data-centre capacity will require about $105 billion of imported equipment, close to 2.5 per cent of GDP. Chips account for the bulk of it, and the bank urges a staggered expansion funded by foreign capital.