Google TPUv7 outperforms NVIDIA GPUs in agentic inference benchmarks
Google’s TPUv7 chip has demonstrated competitive or superior performance per dollar compared to NVIDIA’s GPUs in agentic inference workloads, according to benchmarks released by SemiAnalysis. The results cover four models—Muse Spark Open, GLM-5.2, Qwen3.5 397B, and Kimi K3—using the AgentX benchmark, which simulates multi-turn, long-context, and sub-agent workloads. TPUv7 achieved up to 50%…
1 source
Key points
- TPUv7 outperforms NVIDIA B200/B300 and MI355X by 1.2x–2.7x tokens per dollar in GLM-5.2 AgentX benchmarks at FP8
- Qwen3.5 397B on TPUv7 FP8 beats NVIDIA GB300 NVL72 FP4 by 1.67x in performance per dollar, per SemiAnalysis
- NVIDIA’s FP4 GPUs still lead in some cases but require accuracy trade-offs, while TPUv7’s FP8 results are now competitive
Model pages: Muse Spark Open → · GLM 5.3 →
Read the original at SemiAnalysis · by Bryan ShanOpen source ↗
Topics · follow one to build your own front page
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- AI hardware stocks fall 6% to 21% as midterm results raise construction delays · 12 src
- MetAI helps TSMC and Micron cut chip plant costs with digital twin tech · 1 src
- Nvidia shares fall 8% amid AI spending fears ahead of Nov. 17 earnings · 14 src
- Meta and NewView Oklahoma give 120 veterans Ray‑Ban Meta AI glasses · 1 src
- Supermicro Super AI Station review: convertible DGX Station workstation tested · 1 src
Comments
via GitHub Discussions