Skip to content
Trending storyDeveloping

SemiAnalysis publishes AgentX inference benchmarks for Google TPUv7

2 articles2 sourcessince Nov 5Last article 1h ago ·

Overview

AISummary of 2 articles

SemiAnalysis publishes its first official long-context, multi-turn agentic inference benchmark results for Google's TPUv7, covering four models.

The firm says TPUv7 is either competitive on performance per dollar for external customers or delivers the best state-of-the-art result. One cited result is Muse Spark Open on SGLang AgentX with NVFP4 weights, where SemiAnalysis describes TPU performance per dollar as competitive even though the architecture was co-designed for GPUs.

SemiAnalysis also reports that GB300 FP4 leads on performance per dollar at most speeds for GLM-5.3, and that NVIDIA's counter submissions are still preview results.

Written by AI from the articles below · updated Nov 5, 3:15 PM ET

Check the sources:

Latest developmentNov 5, 2:03 PM ET
SemiAnalysis reports first AgentX inference benchmark results for Google TPUv7

Developments

2 developments

  1. Nov 5, 2:30 PM ET · 1 article
    vLLM and Inferact push vLLM performance on Google TPUs with Kimi K3 megakernel
    vLLM: vLLM and Inferact push vLLM performance on Google TPUs with Kimi K3 megakernel
  2. Nov 5, 2:03 PM ET · 1 article
    SemiAnalysis publishes AgentX inference benchmark results for Google TPUv7
    SemiAnalysis: SemiAnalysis reports first AgentX inference benchmark results for Google TPUv7

Article timeline

The articles in this story. Times are ET.

Nov 5
  1. vLLMOfficial
    vLLM and Inferact push vLLM performance on Google TPUs with Kimi K3 megakernel

    AIvLLM project thanks SemiAnalysis for highlighting its TPU work, including the Kimi K3 megakernel and vLLM TorchTPU powering AgentX results across GLM-5.2, Qwen3.5 397B, and Kimi K3. The post credits Inferact and Google Cloud for pushing vLLM performance on TPU and points readers to SemiAnalysis's write-up on TPU inference.

  2. SemiAnalysisBlog
    SemiAnalysis reports first AgentX inference benchmark results for Google TPUv7

    AISemiAnalysis says it published the first official long-context, multi-turn agentic inference benchmark results on Google's TPUv7 across four models. It reports TPUv7 is competitive on performance per dollar or leads on several FP8 comparisons, while GB300 FP4 leads on perf per dollar at most speeds for GLM-5.3 and NVIDIA's counter submissions are still preview results.

Heat Trend

Not enough continuous observations to show a trend yet.