We’re excited to add support for Stable Diffusion 1.5 INT8 quantization to the Procyon AI Image Generation Benchmark. Image Generation is one of the more demanding subsets of Generative AI, requiring powerful modern hardware to run locally.
With the addition of INT8, the Image Generation Benchmark now supports NPUs found in power efficient devices. This makes Procyon the first available benchmark that has been standardized for measuring and comparing the AI image generation performance of NPUs, iGPUs, and discrete GPUs.
Precision and AI Quality
AI quality is a measurement of an AI workload’s result, such as how accurately a computer vision task identifies objects or the level of detail in a generated image. Changes in the AI model, its precision and the Inference engine can all affect quality.
The following comparison shows how changing AI model precision affects the quality of a generated when using Stable Diffusion 1.5 in the Image Generation Benchmark.
Images from the Procyon AI Image Generation Benchmark using Stable Diffusion 1.5 INT8 and FP16 quantizations. They are generated from the same image generation seed and ran using Intel OpenVINO on Intel hardware.
While the higher-precision FP16 image contains more detail, it requires more powerful hardware and takes longer to generate. Using INT8 allows for faster execution times, a smaller memory footprint, and support for a wider range of hardware, such as modern NPUs.
The INT8 image generation update is a free update for all Procyon customers with a valid annual license for Procyon AI Inference Benchmarks. You only need to update to the latest version of Procyon to access it.
