| Kyuu said: But in layman's terms, what purpose was in the PS5 Pro including a ML hardware accelerator and exclusively supporting PSSR if all of this can be done to arguably a better effect on older/weaker hardware with less features, no dedicated ML block, tensor cores, etc? What is it that DLSS2 or PSSR did that made them better? I'm assuming a lower performance penalty but can you confirm this? What's the point of PSSR? You reckon it will be replaced with FSR4 or will both models evolve separately? Are you able to detect any real world differences between INT8 and FP8 FSR4 yet? I apologize for spamming questions, but my curiosity is nagging on me. |
So there is a lot going on here that influences decisions like this.
To start, you can run ML workloads on regular general-purposed/graphics-focused cores. Tensor Cores/Matrix Cores are just specialized cores to accelerate parallel low-precision workloads, which deep-learning falls under. The problem is that you'll get a bigger performance penalty than when you have specialized hardware, as 1. you are utilizing resources that could be used more efficiently to render the raster pipeline (not an issue if you have specialized cores for this) and 2. you are likely not natively computing but rather casting into a more precise primitive which is more computationally expensive and memory hungry, if you don't have native ALU support for the less precise data-primitive (i.e no FP8 support until RDNA4 means FP8 needs to be computed using FP16 arithmetic units.)
From what I gather FSR4 in its final version is a sort of hybrid architecture that incorporates features of ViTs and CNNs. There are many ways you can hybridize ViTS and CNNs at many steps of the model training process, so I don't know what they specifically did (and people are mostly speculating anyway), but for at least the ViT component you do see major benefits from FP8 precision. All we know about this leaked "FSR4" is that inference is done with INT8, which might mean it is a pure CNN model, a quantization of the ViT-CNN hybrid model, one that has transfer learned or distilled from a better ViT one, or it can just be an old model they had before they moved to a VIT-CNN hybrid, with that move mainly being to sell RDNA4 GPUs with an exclusive feature.
As far as the final result is concerned, it is hard to say. Somebody will have to do an actual experiment using an RDNA4 GPU and released FSR4. It's likely better than the (assumed to be non-lite DLSS 3) implementation in the Switch 2 version of the game, because it is giving a similar result at lower internal and target resolutions when compared to Switch 2 handheld mode. I notice that text is cleaner and the image is sharper on the Steam Deck using FSR4 Performance, but the Switch 2 image is a bit more stable in motion and there is less aliasing overall on Switch 2.
PSSR probably could work on the base PS5, just with a much higher performance impact that might not make it worth it compared to FSR 2/3 in most cases. Still, I think you probably could get better results than FSR2/3 with some lightweight ML model on base PS5.
There is also the matter of the fact that as premier models get better (for architectural reasons), they can help smaller models get better (again through transfer learning and distillation) and it could be possible that hardware which is weaker can utilize some very lightweight, but good DL models as time goes on when it couldn't in the past.
Last edited by sc94597 - on 17 September 2025
