![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
- Joined
- Aug 15, 2005
- Messages
- 9,256
I'm sure you know this, but for the sake of conversation: Tensor cores don't mean DLSS is free.
I think the general point is that maybe a relatively more intensive algorithm can be run, independent of the regular shader engines, etc. And if the right balance is achieved, the GPU cores are free to render the game. and the tensor cores more/less handle the scaling. And the end result is a good image with good performance.
However, DLSS can still be "slow". And therefore hold up the throughput of the final frames. We saw that when DLSS 4.5 released. the "L" preset in particular, is on average about 10 frames slower than DLSS 4 K preset. And it was close to to 20 frames slower, than DLSS 3.
Switch 2 doesn't have many Tensor cores. And they are a relatively older version, with less optimizations. Reportedly, even though Switch 2 is using a "lite" version of DLSS----its still not exactly fast to run it. Some games/developers have chosen to skip it. It sounds like it may be a situation where even though DLSS is offloaded to the tensor cores----its such a straing on those, it can hold up the final frame output in a detrimental way.
I think over time, probably two things will happen, which will make DLSS a better option:
1. Nvidia is able to work with Nintendo to improve the algorithm
2. Developers get more time with Switch 2 dev kits, and gain a better understanding of how to customize and implement DLSS into their Switch 2 games.
------------------
Everyone better understands ML derived code nowadays. So, even the shader based stuff has improved drastically. So much that, we have a user driven project which has ported DLSS3 Preset E, to run on shader code with relatively dated instructions. It doesn't suck. It doesn't look quite as good as the leaked INT8 FSR4. But, it runs faster.
Playstation's new QSSR is running on PS5's GPUs as shader/compute code. Base PS5 GPUs don't have features normally associated with running ML derived code. Yet, they were able to make something which looks great. You're right, being purely shader/compute based, it can potentially take away from the target render speed of the game. AMD is generally very good at asynchronous compute. So, it doesn't seem to be a huge issue. Depending upon the game, testing so far has shown QSSR framerate performance to range from a small hit of maybe 5 frames or in the case of Ghosts of Tsushima----slightly better framerate than FSR3 and/or custom scalers.
AMD has a new FSR 4 "Lite" which is meant for APUs. I'm sure it doesn't look as good as FSR4 on a 9070 XT. And we only have a few seconds of comparions, of a couple of scenes, from an official AMD youtube video.
But, what they show is CLEARLY better than FSR3. In the case of how it resolves the flowing river, its the difference between a totally busted looking image and-----hey that's a pretty normal looking videogame river. and they claim it runs faster than FSR3. They've got some talent working on this stuff, it seems!
That said, I still have to go back to No Man's Sky and their use of FSR 2 on Switch 1: it looks stupidly good. So, true talent and dedication to customize the code rather than pasting it in as general use code library, seems to make all the difference.
Compare all this to the shader based "DLSS" which was featured in Control for a few months, until DLSS 2.0 released and actually used the Tensor cores. It looked fine if you stood still. Looked like artifacting crap, when you started moving. It even looked somehow like an interlaced image. It was pretty interesting and we have come a long way! Epic's own TSR is hardware agnostic and looks pretty dang good, etc.
Last edited:
