Google Expands 8th-Gen TPU Lineup with AMD-Powered 8tx and Dual-Chip 8i Max

Google Cloud has officially expanded its eighth-generation TPU portfolio with the introduction of TPU 8tx and TPU 8i Max, broadening the lineup to four distinct systems tailored for large-scale training (8t), low-latency inference (8i), confidential computing (8tx), and large-context processing (8i Max). Focusing on security, TPU 8tx marks a confirmed partnership with AMD, employing custom AMD x86 host CPUs. Unlike the 8t and 8i variants that utilize Google's in-house Arm-based Axion CPUs, TPU 8tx leverages AMD silicon alongside cryptographic attestation and memory isolation to deliver end-to-end hardware confidential computing across the host CPU, PCIe interconnects, and TPU memory to safeguard model weights and enterprise data during inference. Maintaining 12.6 PFLOPS FP4 compute performance and 216GB HBM, TPU 8tx also integrates the Boardfly low-latency network topology, reducing long-distance chip communications by over 50%. Google clarified that AMD's confirmed involvement currently centers on host CPUs, leaving rumors of broader collaboration on tenth-generation packages unverified, while its custom TPU agreement with Broadcom remains active through up to 2031. To address the memory capacity and throughput demands of multi-step AI agents, TPU 8i Max integrates two TPU 8i chips into a single package. The design features 384GB of HBM (a 33% increase over 8i), 17.2TB/s memory bandwidth, and 768MB of on-chip SRAM, doubling SRAM and bandwidth to raise per-rack compute density and sustain agent throughput. On the software and orchestration front, Google partnered with Meta and industry peers on TorchTPU to enable native execution of PyTorch and Hugging Face models, while previewing open-source TPU agent toolkits compatible with developer environments like Cursor and Claude Code. The new Agent Compute Engine pauses idle sessions during external tool calls to conserve power, and TPU Sync introduces a tiered memory architecture to cut KV Cache transfer latency and improve time-to-first-token. Additionally, Google Kubernetes Engine introduced flexible ComputeClasses for mixed workloads across TPUs and GPUs.

Key takeaways

• Google introduced TPU 8tx, integrating custom AMD x86 host processors to enable hardware-level confidential computing from CPU to TPU memory.
• TPU 8i Max packages two 8i chips into a single package with 384GB HBM and 17.2TB/s bandwidth, optimized for memory-heavy AI agent workloads.
• Google's 8th-gen TPU lineup now comprises four targeted systems covering training, low-latency inference, confidential computing, and high-context throughput.
• Software additions like TorchTPU, Agent Compute Engine, and TPU Sync were unveiled to simplify deployment and reduce cache and idle overhead.