AI Generation Benchmark for GPUs

UL, the company behind 3DMark software suite for GPU and CPU benchmarking, is now launching its first AI generation benchmark for GPUs.

Unfortunately, this benchmark is not part of the 3DMark (consumer) suite, which means it is targeted towards commercial businesses and large customers who are willing to pay extra bucks. The benchmark costs $5000, and it’s available starting from yesterday.

The AI Image Generation Benchmark in the Procyon suite is designed to measure performance of high-end hardware in a fair and responsible way. That is achieved by using one of the most popular and open source generative AI models, Stable Diffusion.

The benchmarks offer two workloads based on the latest XL model for high-end GPUs. In case you don’t know, this model is designed for a higher resolution of 1024×1024, while mid-range GPUs should stick to the older SB 1.5 model, which generates 512×512 images.

To keep benchmark fair and provide optimal performance for any modern AI acceleration hardware, it supports the most important APIs such as:

• ONNX runtime with DirectML (mainly for Radeon GPUs)
• NVIDIA TensorRT (for NVIDIA RTX GPUs)
• Intel OpenVINO (for Intel GPUs)

This is how it looks:

AI Image Generation Benchmark, Source: UL

Since this is not a consumer-oriented test, nor is anything even remotely affordable, it’s doubtful it will become something of great importance among GPU enthusiasts. However, we look forward to reading from tech review sites that have access to many GPUs who might give it a try.

Source: UL