TheRedRiders
Member
Deep Learning Super Soap
That's RTX Blur, which realistically blurs textures depending on the weather and lighting /s
Last edited:
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Nvidia's DLSS 2 vs. AMD's FSR 2: The Ultimate Analysis
- Thread starter Leonidas
- Start date
That's RTX Blur, which realistically blurs textures depending on the weather and lighting /s
analog_future
Banned
Immortals of Aveum: DLSS Performance vs. FSR
FSR Performane mode is not good, but FSR 2.2 Quality mode is fantastic and super close to DLSS outside of a few fringe scenarios that almost no one would notice in normal play anyway.
SolidQ
Member
In real scenario, you won't notice. You can notice only sharp or soap picture.
Bojji
Member
Many games have DLSS sharpen slider, this one don't. Looks like FSR2 has sharpening by default, other than that picture quality is worse with FSR.
SolidQ
Member
in Ghostwire FSR2 was superior implementation, better than Native, in my own test
alucard0712_rus
Member
Show any game where DLSS Quality (at 1440p and above) is looking worse than native with TAA? Especially in motion.
Leonidas
Member
FSR 2.2 Quality at 1440p is not fantastic and still not close to DLSS Quality at 1440p.
FSR 2.2 Quality at 1440p renders at an even lower resolution than 4K Performance.
DLSS2 still holds up well at 1440p Quality.
Zathalus
Member
Yeah because FSR has a sharpening filter applied and DLSS does not. When it comes to literally everything else DLSS is superior.Deep Learning Super Soap
The Cockatrice
Member
Deep Learning Super Soap
Immortals of Aveum: DLSS Performance vs. FSR
You can apply a sharp filter and fix it in 5 seconds regarding your picture
SolidQ
. How are you going to fix the issues FSR brings? Put soap in your eyes? AMD fans..fucking lol
Last edited:
manfestival
Member
analog_future
Banned
I should've clarified that FSR 2.2 Quality at 4k is fantastic. I would agree that I wouldn't use it for lower resolutions/settings.
winjer
Gold Member
AMD was so focused on trying to catch up to nvidia with frame generation, the forgot they still had to catch up to nvidia in temporal upscaling.
Unknown Soldier
Member
They can't really catch up on the upscaling because they don't use AI/ML. No mathematical algorithm will ever be able to match the use of AI in the upscaler.
winjer
Gold Member
But they can use AI for an upscaler. Even if it's just DP4A like with XeSS.
For some unknown, stupid reason, they refuse to use it.
Bojji
Member
Yeah I don't fucking get it, they can keep current version for all GPUs and have higher quality version for newer ones using DP4A. Thos would make Radeon cards look much better.
Gaiff
SBI’s Resident Gaslighter
Yes. It uses ML on Intel Arc hardware. XeSS uses the XMX units to upscalge using Int 8. On older Intel integrated GPUs, it uses DPa4. On Non-intel GPUs, it uses an SM 6.4 path. Last we checked though, XeSS performed quite a bit better on Intel Arc GPUs than on NVIDIA or AMD thanks to using dedicated hardware to reduce rendering time.
Not sure if it has changed much.
Last edited:
Might be to keep as close to dlss performance wise at the same presets, they might think people probably don't notice the image quality differences but if their ai/ml upscaler runs 10fps less than dlss at the same base res/preset that data is easier to show/notice in online benchmarks etc.
winjer
Gold Member
Yes, there would be a drop in performance using DP4A on RDNA2 for AI. But it would be worth it for the better image quality.
On RDNA3, with WMMA support, it would not hurt performance that much.
Gaiff
SBI’s Resident Gaslighter
The DP4a pass seems to run noticeably worse than FSR or DLSS at the same quality in a lot of games though, especially on AMD hardware. It's not so bad on NVIDIA hardware.
winjer
Gold Member
That's because DP4A is just an extension for Matrix operations done in the shaders. So when doing AI, it will meant some shaders can't be used for normal compute operations.
While on the Arch and Nvidia, there are dedicated Tensor Units to deal with Matrix operations.
Gaiff
SBI’s Resident Gaslighter
And you just answered your own question, did you not?
Edit: Also, wait, no. It uses the shader model 6.4 path on non-Intel GPUs. DP4a is only used on Intel integrated GPUs. Has this changed with 1.2?
Last edited:
winjer
Gold Member
I didn't have any question. I was just elaborating on your statement.
XeSS will run on XMX cores for newer Intel CPUs that have these cores.
It will run on DP4A on GPUs that support it, like RDNA2 and later. And Pascal and later.
For other GPUs, it will use SM 6.4
Gaiff
SBI’s Resident Gaslighter
Strange because it was my understanding that DP4a was strictly for Intel integrated GPUs (which I found really bizarre).
I know DF said there were three kernels that worked as you said but there was a separation between DP4a and SM 6.4 support. Maybe it was a misunderstanding or their part or it got updated? Regardless, I never thought it made a lot of sense to reserve the DP4a path strictly for Integrated Intel GPUs when more recent GPUs also support it.
winjer
Gold Member
DP4A path was always intended as a secondary path for all GPUs without XMX units.
Gaiff
SBI’s Resident Gaslighter
Okay, so I looked at the documentation and you should be correct.
An HLSL-based cross-vendor implementation that runs on any GPU supporting SM 6.4. Hardware acceleration for DP4a or equivalent is recommended.
It says "any" GPU supporting SM 6.4 which dates from times immemorial and that hardware acceleration for DP4a is recommended and as you noted, this has been available since Pascal and RDNA2 for NVIDIA and AMD respectively. Wouldn't make a single bit of sense to exclude the DP4a path from GPUs that support it and strictly leave it for Intel integrated GPUs. DF probably misunderstood Intel or Intel wasn't clear. Or maybe this got updated but I doubt it. Would have been stupid.
winjer
Gold Member
Truth be told, DF is far from being the most accurate source for tech news and info.
Gaiff
SBI’s Resident Gaslighter
True but they had a direct channel to Intel for that one. Intel sponsored the XeSS video so DF was actually in contact with them. They're one of the few outlets that can consistently get answers straight from the horse's mouth which is incredibly valuable.
winjer
Gold Member
And despite having a direct contact with Intel, they still screw up.
That's why I said that DF is far from being the most accurate source for tech info.
Gaiff
SBI’s Resident Gaslighter
Yep, but what can you do? It's that or nothing. At least we can cross-reference their reports with what's publicly available.
But yeah some of the shit they get wrong makes me scratch my head sometimes lol.
Last edited:
winjer
Gold Member
Plenty of tech sites and channels that do a much better job than DF.
Gaiff
SBI’s Resident Gaslighter
I can't count that many to be honest. GN and HU for instance are quite a bit better for analyzing. As is de8auer and a handful of youtubers but their pieces are often not as in-depth and more generalized. They also tend to focus much more on hardware than software and seldom get backing or interviews from the devs. Most are also PC-centric so there isn't that much to go by for console coverage.
DF got to sit down with Bluepoint, Nixxes, Intel, CDPR, Turn 10, and a bunch of others and ask them questions that interest us.
If you do have more that I might not be aware of, send them my way and I'll give them a gander.
winjer
Gold Member
GN and HU are first hardware sites. Not so much centered on games. But they do a lot better coverage for benchmarks and tech than DF.
Techpowerup also has specific performance game analysis, including upscalers comparisons.
Then there is Chips and Cheese, probably the most in-depth coverage of hardware and how the software runs on it.
For comparing some games, DF still does a decent job. But to talk about tech, they are lacking.
Gaiff
SBI’s Resident Gaslighter
Yeah, I regularly have been following those for a few years. Chips and Cheese more recently. They're excellent for in-depth coverage but they have a very "developer" like approach to their conclusion. As in, it seems that they can never see developers screwing up.
For instance, they said that there was nothing wrong with Starfield's performance on NVIDIA hardware when typically, AMD equivalent parts crushed them by like 35%. Lo and behold, with that big beta patch, NVIDIA is now much closer to AMD. AMD still runs better but by like 10-15%, not the outrageous performance differential there was before. Now I think the 4090 is substantially faster than the 7900 XTX whereas before, they were almost equal.
Last edited:
winjer
Gold Member
Yeah, thy underestimated Bethesda's incompetence.
Still, they don't have access to the source code, so they can only see how instructions are passed in the hardware.
JimboJones
Member
I don't think this is true, at least not entirely, FSR2 as a technology is definitely better than most solutions used back then, especially PS4s checkerboard upscaling.
I think what's happening is developers are pushing resolutions to way more extremes.
If your think FSR2 is bad at those extremes the results from something like ps4 checkerboard at those extremes would probably look even worse.
Shake Your Rump
Member
I wonder if this is due to resolution and not the implementation? Checkerboard on PS4 would use a 1440 or higher source. Games are now rendering as low as 720, so it's going to look much worse even with a superior upscaler.
Gaiff
SBI’s Resident Gaslighter
No, they weren't. Checkerboard was just typically operating at a much higher internal resolution because the target fps was often 30 anyway. Typically, checkerboard rendering used 1920x2160 to upscale to 4K. That's much, much higher than games using FSR2 to upscale to 4K from a 1600x900 resolution.
It was always supported through SM 6.4.
"Yes, we require inference capabilities but matrix acceleration is not the only form of inference capability that is available on GPUs. If you go all the way back to- I think Skylake- we had dot product acceleration, which is DP 4.4 – there’s various names for it. Nvidia has had this I think, since Turing and AMD has this now on RDNA2. So even without Matrix acceleration you can go quite far. It might not be as fast as matrix acceleration, but certainly meets the objective. And as I said, the objective is to maintain the fidelity of your render and achieve smooth frame rates. So, when it comes to older models, on older internal GPUs, we've had dot product acceleration (DP4a) for a while now.
Microsoft has enabled this through Shader Model 6.4 and above and on all these platforms XeSS will work."
The Intel XeSS Interview with Principal Engineer Karthik Vaidyanathan
Today we will be joined by Intel Principal Engineer Karthik Vaidyanathan and talk extensively about Intel's upcoming neural supersampling technology, XeSS, and that very impressive demo. Karthik was instrumental to XeSS development and will be fielding the community's questions about the...
wccftech.com
However I believe that integrated Intel GPUs use a different DP4a path that bypasses SM 6.4, so that's probably where the confusion came from.