AMD

"There are better ways to use AI in games than image processing"...David Wang (AMD)

With generation RDNA 3, AMD expanded AI acceleration support with WMMA matrix multiplication. In the interview that David Wang and Rick Bergman gave to the website 4gamer, this subject – which has been much talked about lately, but at the same time somewhat controversial – came up. AMD's implementation differs significantly from Nvidia's approach, which supplemented the core with separate hardware inference accelerators (Tensor Cores). The question directed at Wang was what he thought about GPU inference acceleration.

“Even the latest generation of CPUs can handle inference processing quickly enough. About 95% of the world's AI inference workload is done on CPUs. This also applies to data centers and our Epyc server processors, which are part of many AI processing systems. .

Where I see room for GPU AI processing is in machine learning (training) AI.

I believe that tasks using the inference accelerator installed on v[herním] GPUs should not be limited to the image processing represented by Nvidia's DLSS. Check out FidelityFX SuperResolution (FSR), one of FidelityFX's many features. High-resolution anti-aliasing and FSR processing implemented without the use of an inference accelerator provides performance and quality that can fully compete with Nvidia's DLSS. The reason Nvidia tries to use AI even for tasks that can be done without AI is because Nvidia has built many inference accelerators into the GPU. In order to be able to use them effectively, she was probably looking for the kind of load that could support [paralelně] make use of many inference accelerators. That's the GPU strategy for gaming, which is great, but I don't think we should go the same route.

Our goal is to implement features that users want and need that benefit the consumer segment. Otherwise, users pay for features they won't use. Therefore, I believe that gaming GPUs inference accelerators should be used for purposes that make games more advanced and fun.”

Intel XeSS technology, competitor of Nvidia DLSS and AMD FSR, which, like Nvidia's solution, uses an artificial intelligence accelerator (specifically, inference) can also be mentioned here for a complete overview. Despite this, it achieves greater performance drops and lower visual quality than AMD solutions.

Source: 4gamer