Many MindedGoogle adds confidential-compute and double-chip TPU systems to its lineup
Google has added two systems to its eighth-generation TPU lineup: the TPU 8tx and the TPU 8i Max. Google says this is the first time it has shipped four TPU systems within a single year. Both new variants reuse existing silicon. They are new system configurations with different packaging, networking and memory setups, not a new chip generation.
The TPU 8tx is aimed at enterprises and government agencies handling confidential data. It keeps data and model weights protected and isolated while they are being processed. It uses the Boardfly network, built for real-time inference, and is meant to deliver up to twice the latency reduction versus the TPU 8t. Google says a single system can handle fine-tuning, distillation of large models into smaller ones, and reinforcement-learning training of agents without moving sensitive data between clusters. The TPU 8i Max puts two TPU 8i chips on one package, giving double the compute, memory and memory bandwidth in the same rack format.
Why it matters: a modular approach that Google says delivers new hardware variants in months rather than years could make specialized AI hardware arrive faster. Confidential computing also matters for access. Organizations holding sensitive data, such as government agencies, could use AI on it without exposing it, which widens who can use these tools. Google also released TorchTPU as open source so PyTorch code can run on TPUs with minimal changes. It supports the vLLM and SGLang inference frameworks, and counts nearly 100 external TPU customers.
Caveats: the performance figures are Google's own statements. The source gives no pricing, availability dates or independent benchmarks. Nvidia already offers a comparable confidential computing feature for its GPUs, and Amazon, Microsoft and OpenAI are building their own accelerators. Whether any of this lowers costs for users remains to be seen.
Source: The Decoder