To provide meaningful and realistic image generation benchmarks, Procyon aligns each workload with the hardware class best suited to run it.  

Modern NPUs are designed for efficient AI inference rather than maximum throughput. As a result, Stable Diffusion 1.5 Light is the most appropriate image generation benchmark for Intel AI Boost, AMD Ryzen AI NPU, Qualcomm Hexagon NPU and Apple Neural Engine (ANE) enabling meaningful measurement of on-device generative AI capabilities. Beyond NPU performance, the benchmark can also showcase platform capabilities such as OpenVINO hybrid NPU-GPU execution. Apple platforms can also be evaluated using CoreML all mode, which is supported in both Stable Diffusion 1.5 Light and Stable Diffusion 1.5. 

For integrated GPUs, Stable Diffusion 1.5 serves as the ideal baseline benchmark. Its moderate compute requirements make it suitable for integrated GPUs while also providing a common workload across Intel, AMD, NVIDIA, and Apple hardware.  

At the high end, SDXL provides a significantly more demanding workload that highlights the capabilities of powerful discrete GPUs through backends such as TensorRT for RTX, ONNX, OpenVINO, and CoreML.  

Vendor-optimized backends generally provide the best balance of performance, memory efficiency, and compatibility because they leverage hardware-specific acceleration features. Where a vendor-specific backend is unavailable, ONNX Runtime with DirectML offers broad compatibility across supported Windows graphics hardware.

For best performance and compatibility, Procyon automatically selects the preferred inference backend for supported hardware. In general, the recommended backend follows the hardware vendor:

  • NVIDIA - TensorRT for RTX
  • Intel - OpenVINO
  • Qualcomm - QNN
  • AMD Ryzen AI - ONNX + Ryzen AI
  • Apple Silicon - Core ML

Note: The ONNX implementation used for Stable Diffusion 1.5 Light for GPUs executes in FP16 precision. Results from this path are therefore not directly comparable with NPU backends that employ mixed-precision execution. 



Hardware Class

Recommended BenchmarkPrecisionPrimary BackendsExplanation
NPU

SD Light 


(512×512, batch 1, 50 steps) 


w8a16Ryzen AI, OpenVINO NPU, QNN, CoreMLLightweight diffusion workload optimized for dedicated AI accelerators and AI PCs.  
Integrated GPU (iGPU)

SD 1.5 


(512×512, batch 4, 100 steps) 


FP16ONNX, OpenVINO, CoreML, TensorRT for RTXProvides a balanced workload with good image quality and broad compatibility across mainstream graphics platforms.
Discrete GPU (dGPU)

SD 1.5 + SDXL 


(1024×1024, batch 1, 100 steps) 


FP16TensorRT for RTX, OpenVINO, ONNX, OpenVINOEnables evaluation of both mainstream and high-end image generation performance, with SDXL showcasing premium GPU capabilities.