InfraSep 24, 2026 · 03:437 sources

TPUv7 Outperforms Nvidia GB200 by 56% with First Kimi K3 Megakernel

From marketbrief.now

Image: marketbrief.now
AI;DR

TPUv7 hardware reached 709 tokens per second during low-concurrency decode for the Kimi K3 model, outperforming the Nvidia GB200 baseline of 450 tokens per second by 56%, according to vLLM maintainers and Inferact. The improvement was driven by a first-of-its-kind TPU megakernel optimization, part of a broader trend toward externalizing software for TPU hardware beyond internal Google environments.

Read the sourceSign in to suggest

Sources

+ 2 more
  1. SUPPORT@sawyermerritt“Next week, @SpaceX will be bringing online 220,000 Nvidia GB300 GPU (worth $15+ billion)”x.com
  2. SUPPORT@sawyermerritt“That's over $100 billion of GPUs combined, and brought online in record time. Over 2 GW of compute.”x.com
#tpu#nvidia#gb200#kimi#vllm#megakernel#elon-musk#chipsMarkdown
Content history (2)
  • 2026-10-04 · Summary · vi · Replaced stored text
    Các nhà duy trì vLLM và Inferact báo cáo rằng phần cứng TPUv7 đã đạt 709 token mỗi giây t…
    Nhóm duy trì vLLM và Inferact cho biết phần cứng TPUv7 đạt tốc độ 709 token mỗi giây tron…
  • 2026-10-04 · Title · vi · Replaced stored text
    TPUv7 vượt trội hơn Nvidia GB200 56% với Megakernel Kimi K3 đầu tiên
    TPUv7 vượt trội hơn Nvidia GB200 tới 56% nhờ megakernel Kimi K3 đầu tiên

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today