Skip to content
The Markets Desk — Week 44 · Wed 28 Oct 2026
Data / Chips · Accelerators

How fast the chips actually serve a model.

How fast a system serves gpt-oss-120b
as of 2026-09-16 · checked 2026-10-26
1,039,900 tokens a second1m

MLCommons · the Federal Register · biannual · continuous · history from the source back to 2 Dec 2024 · since 21 Sept 2026

we gathered it

AMD 9x8xMI355X_2xEPYC_9575FAMD 8xMI300X_2xEPYC_9575F

AMD 9x8xMI355X_2xEPYC_9575F serves 1,039,900 tokens a second on MLPerf's offline test.

Nothing has landed here yet.

  • 2026-10-28SK Hynix

    2026-Q3 numbers

  • 2026-10-28KLA Corporation

    2026-Q3 numbers

  • 2026-10-28Silicon Laboratories

    2026-Q4 numbers

  • 2026-10-28FormFactor

    2026-Q3 numbers

  • 2026-10-28Wolfspeed

    2026-Q3 numbers

  • 2026-10-28Amphenol

    2026-Q3 numbers

  • 2026-10-28Impinj

    2026-Q3 numbers

  • 2026-10-28United Microelectronics

    2026-Q3 numbers

  • 2026-10-29Allegro MicroSystems

    FY2027-Q2 numbers

  • 2026-10-29Entegris

    2026-Q3 numbers

  • 2026-10-29Monolithic Power Systems

    2026-Q3 numbers

  • 2026-10-29STMicroelectronics

    2026-Q3 numbers

What happened

on this beat, newest first
The sources behind this tracker
SourceCadenceRows this monthNewest figureLast fetchMisses
MLCommonsclaimed-vs-delivered throughput per vendorbiannual02026-09-162026-10-260
the Federal Registerthe accelerator chokepoint calendar — rule effective dates and entity-list additions; forward events a lean can be sealed oncontinuous12026-11-202026-10-280

Take it with you

Coming in Q4: embed this chart · tell me when it moves