Google has just expanded its TPU v8 family by introducing two new variants for agentic workloads, unlocking confidential inference and small-scale training with the TPU 8tx, while the TPU 8i Max attempts to solve the memory wall problem. An overview of Google’s TPU 8t and TPU 8i chips Google’s TPU 8t is housed within the “Zebra package” that physically integrates a central Logic chiplet, a dedicated SerDes chiplet, and multiple high-bandwidth memory stacks. TPU 8t compute and processing units As we noted in a dedicated post in April, Google’s TPU 8t is primarily meant to train AI models. The core […]