AMD

AMD has the one thing every NVIDIA challenger needs and almost none of them have: a track record of taking market share from a dominant incumbent, which it did to Intel in server CPUs over the previous decade.

Instinct accelerators compete credibly on memory capacity and bandwidth, which matters enormously for inference on large models. The gap has always been software, and ROCm is the long bet: an open stack that customers can fork rather than a proprietary one they must rent.

AMD is also unusually straight with its numbers. It quotes both dense and sparse throughput, which makes its figures less flattering and more useful than a headline that only gives you the larger one.

Silicon from AMD

8 chips on record, newest first.

  1. 2026

    AMD Instinct MI350P

    TSMC 3nm | 6nm FinFET PCIe Gen5 card, CDNA 4, air-cooled, 600 W

    MXFP4: 4.6 PFLOPs

    Absolute
  2. 2026

    AMD Instinct MI455X

    TSMC 2nm | 3nm FinFET CDNA 5 GPU (MI400 Series) in the Helios rack, 72 GPUs per rack

    Peak OCP MXFP4 Performance: 40.3 PFLOPs

    Absolute
  3. 2025

    AMD Instinct MI430X

    Not stated by AMD MI400 Series GPU for HPC and sovereign AI, keeps hardware FP64 (288 TFLOPS)

    9.2 PFLOPs FP4

    Absolute
  4. 2025

    AMD Instinct MI355X

    TSMC 3nm | 6nm FinFET CDNA 4 GPU, OAM, liquid-cooled

    MXFP4: 10.1 PFLOPs

    Absolute
  5. 2024

    AMD Instinct MI350X

    TSMC 3nm | 6nm FinFET CDNA 4 GPU, OAM, air-cooled

    MXFP4: 9.2 PFLOPs

    Absolute
  6. 2024

    AMD Instinct MI325X

    TSMC 5nm | 6nm FinFET CDNA 3 GPU, OAM

    Peak FP8: 2.61 PFLOPs

    Absolute
  7. 2023

    AMD Instinct MI300X

    TSMC 5nm | 6nm FinFET CDNA 3 GPU, OAM, 153bn transistors

    Peak FP8 Performance (E5M2, E4M3): 2.61 PFLOPs

    Absolute
  8. 2023

    AMD Instinct MI300A

    TSMC 5nm | 6nm FinFET APU, 24 Zen 4 CPU cores plus CDNA 3 GPU on one package

    Peak FP8: 1.96 PFLOPs

    Absolute

Compare against every other vendor in the silicon tracker

Our coverage of AMD