A benchmark version number is specific to a test. Benchmark version numbers change rarely and only when absolutely necessary to accommodate changes in third-party applications or bug fixes.

Procyon AI Computer Vision 2.0 Benchmark v1.2.612

Sep 18, 2026

  • NVIDIA TensorRT-RTX backend 
    • Added ARM64 target for native TensorRT-RTX backend 
    • Updated CUDA SDK to v13.4.1 for both x64 and ARM64 targets  
  • Qualcomm SNPE backend 
    • Re-converted models and updated runtime to QAIRT SDK v2.48.40 
  • WinML+ QNN EP backend 
    • Re-converted models and updated JIT model compilation dependencies to QAIRT SDK v2.48.20 
  • For the SNPE and WinML-QNN backends on Qualcomm, SAM2.1 Encoder and BLIP Encoder models now have an additional ONNX optimizer pass in the model generation recipe. This pass replaces Multi-Head Attention blocks with Single-Head Attention blocks for better performance on Qualcomm NPUs. These changes improve performance on X2 Elite by over 30% and by 25% on X Plus (1st Generation) in Overall scores in our test sample.
  • Windows ML
    • Added support for standalone ARM64 TensorRT-RTX EP alongside existing standalone EPs
    • Updated the standalone QNN execution provider to v2.4.0.0

Notes

  • Please note that the first run using WinML-QNN may take notably longer than usual as the models need to be compiled. Once compilation is complete, subsequent runs should reuse the compiled models and run faster.
  • For the Computer Vision 2 test, benchmark results must be refreshed for the WinML with TensorRT RTX EP backend using version 1.3. We recently identified and corrected a behavior change in the WinML TensorRT RTX EP backend that caused some NVIDIA GPU configurations to report incorrect ConvNext throughput numbers. As a result, previously collected benchmark data is no longer representative of current performance. Following this correction in version 1.3, the overall score regressed by more than 3% on 50% of the tested NVIDIA systems in our sample, with affected systems averaging a 10.5% decline. To ensure accurate and consistent comparisons, all WinML and TensorRT RTX EP benchmark results should be regenerated with version 1.3 and used as the new baseline for future evaluations.


Procyon AI Computer Vision 2.0 Benchmark v1.1.593

Aug 17, 2026

Updated

  • WinML updated from 1.8.2124 to 2.1.74 
  • Added WinML standalone EP support for all vendors 
  • Updated SNPE models and runtime to SDK 2.47 
  • Updated OpenVINO to 2026.2 
    • Fixed bug where running the test on a specific target hardware creates model caches that would be incorrectly used for other target hardware on the system
    • Enabled driver-based NPU model compilation flag
  • Updated TensorRT RTX to 1.6
    • If the system has the WinML 2.0 compatible NVIDIA RTX EP v0.0.40.0, the workload will use the standalone EP instead as this specific version crashes when running the ESRGAN model. The newer EP v2.30.43.0 released on 29.7.2026 resolves this issue and is used if available

Procyon AI Computer Vision 2.0 Benchmark v1.0.567

May 27, 2026

Updated

  • Added support for local generation of EP Context models for AMD Vitis AI EP. This enables the workload to automatically generate compatible models that benefit from the latest VitisAI EP updates without the need to update the benchmark with every EP release. 
  • Enabled shared memory buffer usage for OpenVINO and QNN Execution Providers. Helps avoid implicit IO memory transfers by utilizing shared memory buffers bringing minor performance improvements 
  • Updated native Intel OpenVINO inference engine to v2026.1 
  • Updated ESRGAN model used in WinML + QNN Execution Provider (EP) path

Procyon AI Computer Vision 2.0 Benchmark v1.0.557

April 17, 2026

Added

  • Added native Ryzen AI path for systems with AMD NPUs based on XDNA 2 or later architectures.

Procyon AI Computer Vision 2.0 Benchmark v1.0.555

April 13, 2026

Fixed

  • Resolved an issue where an incorrect SNPE DLL was packaged with the WinML QNN Execution Provider runtime.
  • Model cache generated for WinML QNN Execution provider use SDK 2.39 to avoid regression in BLIP model caused by SDK 2.40 driver dependencies. 

Added

  • Model caches generated for the WinML OpenVINO Execution Provider now include the execution provider version in the cache name so that caches can be regenerated when a new OpenVINO EP is released. 
  • Improved error logging for WinML workloads when Execution Providers fail to download.  

Procyon AI Computer Vision 2.0 Benchmark v1.0.539

March 30, 2026

This update is the release version of the AI Computer Vision 2 for Windows.

Release Workloads:

  • Image classification with ConvNeXt
  • Image captioning with BLIP
  • Object detection with DETR
  • Image segmentation with SAM2
  • Video upscaling with Real-ERSGAN

Release Execution and Runtime Support

Windows

  • Support for Microsoft Windows ML as a standardized execution path, with the following execution providers
    • DmlExecutionProvider
    • NvTensorRTRTXExecutionProvider
    • OpenVINOExecutionProvider
    • QNNExecutionProvider
    • VitisAIExecutionProvider
  • Native inference paths remain available for comparison
    • NVIDIA TensorRT
    • Intel OpenVINO
    • Qualcomm SNPE
  • Supported accelerators include GPUs and NPUs across major vendors

macOS

  • Core ML is the default execution path on Apple Silicon-based Macs.