Nvidia’s GB300 FP4 submission trails Google’s TPUv7 on cost efficiency
SemiAnalysis’s October 19, 2026 report compares Nvidia’s GB300 FP4 entry with Google’s TPUv7 on an agentic inference benchmark used by external customers. The analysis finds Nvidia’s submission 1.64 times worse in performance per dollar, while Google’s TPUv7 scores between 1.42 and 1.67 times better. Nvidia’s entry runs in FP4 (four‑bit) format, whereas Google’s uses FP8 (eight‑bit), making FP4…
Key points
- Nvidia’s GB300 FP4 submission is 1.64× worse in performance per dollar than Google’s TPUv7 on agentic inference, per SemiAnalysis.
- Google’s TPUv7 achieved 1.42–1.67× better cost efficiency using four chips, while Nvidia used a full 72‑GPU rack.
- GB300 NVL72 offers up to 15 PFLOPS FP4 per GPU (1.1 exaFLOPS per rack) but still trails Google’s FP8‑based TPUv7.
The report notes that the cost‑efficiency gap is most pronounced on medium‑to‑high interactivity tasks typical of agentic AI, where long conversations and multi‑step reasoning drive many inference calls. SemiAnalysis expects further counter‑submissions from Nvidia as it works to close the gap. For investors and enterprises, the findings suggest that Google’s architecture may offer a cheaper scaling path for agentic workloads, especially when starting with a small chip count.
Nvidia’s GB300 FP4 submission still trails Google’s TPUv7 on cost efficiency
cryptobriefing.com · 19 October 2026
Loading the full article…
This text was published by cryptobriefing.com and written by Diego Almada Lopez. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
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 →- Tenstorrent raises $1.5B Series E at $6.2B valuation · 3 src
- Meta brings Prometheus supercluster online in Ohio to train Muse models · 2 src
- Opinion: TSMC seen as AI chip stock with most upside through 2030 · 2 src
- DeepSeek plans to use Huawei Ascend 950DT for pre‑training by end of 2026 · 3 src
- Nvidia says Vera CPU delivers 3.4 TB/s bandwidth and higher single-thread performance · 6 src
Comments
via GitHub Discussions