Google Reportedly Taps AMD for 10th-Gen TPU as 1:1 CPU-to-Accelerator AI Designs Gain Traction
Updated
Updated · Tom's Hardware · Aug 16
Google Reportedly Taps AMD for 10th-Gen TPU as 1:1 CPU-to-Accelerator AI Designs Gain Traction
3 articles · Updated · Tom's Hardware · Aug 16
Summary
SemiAnalysis said Google is working with AMD on a 10th-generation TPU project, potentially marking AMD’s first major role in a custom AI ASIC.
The reported tie-up points less to a standard TPU build than to a CPU-heavier design, with on-package CPU cores aimed at reinforcement learning, reasoning and other general-purpose-intensive AI workloads.
Google has already been adding more host compute around TPUs: TPU 8i systems use one Axion CPU for every two TPUs, versus one Xeon for every four 7th-generation TPUs.
SemiAnalysis said some AI setups may work best at a 1:1 CPU-to-accelerator ratio, making tighter CPU-accelerator integration attractive for both performance and power efficiency.
AMD’s appeal could lie in CPU IP and advanced packaging, including experience from the MI300A hybrid design, though the exact scope of its involvement remains unconfirmed.