01/28/2023 at 7:15 p.m. by Norman Wittkopf – Artificial intelligence and deep learning are on the rise. A current test put different Geforce and Radeon graphics cards to the test.
In terms of artificial intelligence (AI) and machine learning, the text bot Chat-GPT and image software, which creates new art on the basis of collected data using deep learning, have recently attracted attention. While many such applications run externally with a lot of server power behind them, you can also run the same thing on your own PC with the computing power of the existing graphics card using the text-to-image program Stable Diffusion.
The website Tomshardware.com has taken advantage of this with numerous Nvidia, AMD and Intel GPUs and recorded how these perform in benchmarks. In summary, Nvidia graphics cards have the upper hand when it comes to processing, as most software in the area is developed using CUDA and other Nvidia tools, according to the report.
Nonetheless, Stable Diffusion naturally runs on other GPUs as well, although three different projects of the software were used for the test by Tomshardware.com, because none of them consistently works on all graphics cards from the three major manufacturers. Considering options and usability, Nvidia decided to use the Webui version of Automatic 1111, while AMD GPUs were tested with the Shark version of Nod.ai.
Mixed results
Support in terms of performance and operability was the most difficult with Intel’s Arc GPUs, but with Stable Diffusion OpenVINO there were “some very basic functions” here too. Comparability is difficult overall due to the different versions and partly also the use of Linux.
In addition, a fix had to be installed for the performance of the RTX 40 cards, while RDNA 3 GPUs would do very well according to the wording and RDNA 2 GPUs would appear to be rather mediocre. In the meantime, the RX 7900 series could not be tested under Linux in one case and the results also vary depending on the use of the arithmetic units.
In conclusion, it is ultimately at best a snapshot of the performance of and in Stable Diffusion. Frequent project updates, support for various training libraries, and more would change the performance profile, so they intend to revisit it at a later date with possibly more broad-based support.
