The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference

Google's TPUv7 Ironwood has demonstrated strong performance per dollar on long-context, multi-turn agentic inference benchmarks across several frontier models, according to recent SemiAnalysis findings. The TPU results show competitiveness with or superiority over NVIDIA's B200, B300, and GB300 chips in specific FP8 configurations, driven by ecosystem collaborations with tools like SGLang and vLLM. However, NVIDIA's FP4 implementations and optimizations continue to maintain advantages in certain end-to-end performance metrics. Open article →