Cross-platform NPU benchmark

What does your NPU really deliver?

Actual maximum and real-world TOPS compute benchmarking.

Run it with your apps open and closed to see if they are limiting your TOPS, and spot underperforming GPUs and NPUs in multi-card machines.

Live from the results database

Ranked runs 4
This week 4
Chips 3

Latest results

  1. NVIDIA GeForce RTX 3070 Ti 32.0 DirectML · FP16 · Windows · score 3.2
  2. Apple Silicon 4.47 DirectML · FP16 · Windows · score 3.2
  3. Apple M5 Max 57.7 Core ML · A8W8 · macOS · score 3.2
  4. Apple M5 Max 60.9 Core ML · A8W8 · macOS · score 3.2

Top chips this week · score 3.2

  1. Apple M5 Max 59.3 2 runs · median
  2. NVIDIA GeForce RTX 3070 Ti 32.0 1 run · median
  3. Apple Silicon 4.47 1 run · median

Why NPUMark

  • Measured, not marketed.

    NPUMark runs the peak test itself — 2 ops per MAC — and shows what your chip actually reaches. No multipliers, no marketing figures.

  • Every engine, every processor.

    CPU, integrated and discrete GPU, NPU: Core ML and the Apple Neural Engine, MLX, ONNX Runtime with Qualcomm QNN, Intel OpenVINO, AMD Vitis AI (Ryzen AI), DirectML, CUDA, LiteRT and NNAPI on Android. NPUMark finds the fastest path on your machine.

  • Right answers only.

    Every network runs its eval set against full-precision reference outputs and must pass an accuracy gate; a fast wrong answer scores nothing.

  • Real networks, not just peaks.

    Image classification (MobileNetV3, ResNet-50), segmentation (DeepLabV3) and sentence embeddings (Granite Embedding 30M) show how much of the peak real work gets; an on-device LLM (SmolLM2-360M) is measured on request and reported unscored.

  • One score, every platform.

    macOS, Windows (x64 and Arm), Linux (x64 and Arm), iOS and Android — compare a phone, a laptop and a workstation on one scale.

  • Verifiable.

    Every shared result is signed, gets a verification code and a public page; the workload pack is pinned and checked before every run.

  • Workstation-ready.

    Multi-GPU box totals and a large-batch peak for big cards (score 3.2, shown beside the score, never in it).

  • Free.

Measure yours.

Run NPUMark, get a signed result and a public page you can share.