Google may turn to AMD for a CPU-heavy TPU design

Google has reportedly teamed with AMD on one of its 10th-generation TPUs, according to a SemiAnalysis client note cited by Sean. If accurate, the work would mark AMD’s first major role in a custom AI ASIC project and point to a possible shift in Google’s accelerator strategy toward designs that pair TPU compute with on-package CPU cores.

SemiAnalysis suggested AMD’s appeal may come from its CPU IP, advanced packaging expertise, SoIC capabilities, programmable logic or interconnect technology. Google already has extensive TPU design experience and has worked with Broadcom as its silicon design partner, making a conventional TPU assignment less likely.

The CPU angle is tied to reinforcement learning, reasoning and agentic workloads, which can require more general-purpose compute around accelerator operations than conventional LLM training. Google’s TPU 8i systems use one Google Axion CPU for every two TPUs, while servers using Google’s 7th Generation TPUs used one Xeon ‘Emerald Rapid’ processor for every four TPUs.

AMD’s Instinct MI300A shows experience combining x86 and accelerator chiplets in a data center-grade package. The reported project remains unconfirmed, but the potential significance is a CPU-heavy TPU v10 variant optimized for reinforcement learning and agentic models rather than another standard accelerator.