Google Adds TPU 8i Max and TPU 8tx to Its Eighth-Generation TPUs: 8i Max Packs Two Chips for Double the Compute, SRAM and HBM Bandwidth, 8tx Targets Confidential Inference, and TPU Sync Goes Open Source

Background and Pain Points

Agent workloads are changing what inference hardware needs. When many agents read an entire codebase or hours of video at once, memory bandwidth becomes the bottleneck, and regulated enterprises and governments also need sensitive data and model weights to stay protected while they are being processed. In April at Cloud Next, Google launched the eighth-generation TPU 8t for training and TPU 8i for inference; it is now adding two more systems in the same generation to address these two problems. Google says TPUs have nearly 100 customers, including frontier labs, trading firms, retailers and universities.

Architecture Highlights

Google stresses that these are new systems, not new chips. The silicon stays the same; packaging, memory, networking and security change, which lets Google deliver in months. TPU 8i Max puts two 8i chips in one package, doubling compute, SRAM and HBM bandwidth together in the same rack footprint. TPU 8tx uses the 8t chip with 5th Gen AMD EPYC CPU hosts and adds confidential computing in hardware, so data and weights stay encrypted and isolated even while processed in memory. It also uses the low-latency Boardfly network built for real-time reasoning, so teams can run private fine-tuning, distillation and reinforcement learning on one system.

Benchmarks

All figures below come from Google; there are no third-party tests yet. 8i Max has 2x the compute, SRAM and HBM bandwidth of 8i. 8tx has up to 2x lower latency than 8t with a near-zero performance penalty. TPU Sync cuts KV transfer latency and time to first token by more than 90%. GKE dynamic slicing makes job startup up to 5x faster and fault recovery up to 4.5x faster. Google says its vLLM plus TorchTPU stack reached leading price-performance on the InferenceX and AgentX benchmarks, but the blog gives no specific scores. Peak compute, availability dates and pricing for the new systems are not published.

Getting Started

For most developers the software news matters more than the new hardware. TorchTPU, open-sourced last month, runs PyTorch and Hugging Face models on TPUs without code rewrites, and vLLM and SGLang have ready TPU backends. The new preview TPU agent suite plugs into Claude Code, Cursor and Antigravity to handle TPU onboarding, find bottlenecks and tune kernels in plain language. Teams already running TPUs on GKE can try Cloud Pathways and dynamic slicing, both now generally available. Teams that need confidential inference should watch for 8tx availability dates and regions before planning a move.