UL Solutions – best known for the 3DMark benchmarks – is releasing its first challenging AI benchmark on March 25, 2024. The so-called Procyon AI Image Generation Benchmark is primarily intended for standalone graphics cards. Due to the power required, their calculators are well suited to generating images based on text prompts.
Advertisement
For consistent measurement values, the Procyon AI Image Generation Benchmark uses reference images that the GPUs have to render. The whole thing looks similar to the rendering benchmark Cinebench, but works fundamentally differently. No ray tracing rays are calculated here.
The common stable diffusion model serves as the basis in an initial version – UI plans to add further tests in the future. By default, the benchmark on Intel graphics cards like the Arc A770 uses Intel's own OpenVINO toolkit. Nvidia's TensorRT is used in GeForce graphics cards. Since AMD does not offer its own suitable toolkit, UL uses ONNX with Microsoft's DirectML interface for Radeon GPUs.
Anyone who wants to experiment should also be able to exchange the underlying toolkits. Theoretically, those interested should also be able to test how quickly processors can generate AI images – only with longer waiting times and lower points.
Can only be used as an expensive license
However, the UL benchmarks, which run under the Procyon brand, are primarily intended for companies and the press. There are no affordable individual licenses. Private individuals can use Stable Diffusion for image generation and testing.
The AI Image Generation Benchmark is UL's second AI benchmark. The company presented the first as the “AI Computer Vision Benchmark”: It tests the image recognition of CPUs, GPUs and small integrated AI units, for example for identifying animals.
(mma)