• 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.

FidelityFX Super Resolution got me re-excited for current consoles. Nearly a half gen upgrade?

LordOfChaos

Member
Video didn't seem to mention anything about the PS5. Not really sure why I should care what some random site says. I can find random websites to say pretty much whatever I want all over the Internet.

My argument was from someone who stated AMD said so but no one seems to be able to provide this information.

I would hope the PS5 would get the feature but with a custom system like the PS5 it may not be that simple as just a software update.

I remember them saying it was coming to every RDNA family GPU including consoles in the video, the Linus timestamp I added also talked directly about that and why they delayed it to work on the whole family together
 

assurdum

Banned
I remember them saying it was coming to every RDNA family GPU including consoles in the video, the Linus timestamp I added also talked directly about that and why they delayed it to work on the whole family together
It's a software solution as the lately VRS one (which should work as well on ps5). Of course it will work on ps5. Now we need to see if developers will use it or they prefer their own solution.
 

johntown

Banned
I remember them saying it was coming to every RDNA family GPU including consoles in the video, the Linus timestamp I added also talked directly about that and why they delayed it to work on the whole family together
Yeah I recall the every RDNA part but maybe I didn't listen or missed the console part?
 

johntown

Banned
I don't follow you. Of course won't work in the older games without a dedicated coding but doesn't means ps5 can't use it for the new ones.
Yeah new games could be built with support and maybe they are even in the works now? Generally, it won't be a toggle on/off feature.
 

johntown

Banned
Can I ask to you why shouldn't work on ps5?
The PS5 uses a custom AMD chip running on a custom OS. Maybe what the Fidelity FX uses for the machine learning is in use for something else in the system and cannot be used for up-scaling? Just a thought
 

LordOfChaos

Member
Yeah I recall the every RDNA part but maybe I didn't listen or missed the console part?


Consoles were included as part of the aforementioned family, I vividly recall them including the consoles and coming here to talk about it, can't find the exact second but that's exactly what Linus is talking to



zLZMFbO.png
 
Last edited:

DonkeyPunchJr

World’s Biggest Weeb
Bovine feces.
There is no "tensor core dedicated to DLSS".
Nor is there anything magical about "tensor cores" themselves, essentially, it is "a bunch of FP ops".

And the operation used to improve what otherwise is 90%+ TAA is called "neural network inference". GPUs are pretty darn good at it out of the box.
A. Tensor cores are much smaller and faster than CUDA cores at the cost of reduced precision. 119 Tensor FP16 TFLOPS for the 3080. Up to double that for sparse matrices. So that is a pretty huge performance boost for workloads that are amenable to it (e.g. inferencing)

B. Also has the benefit of using a separate, dedicated piece of silicon rather than using the standard CUDA cores/Stream Processors and thus taking some performance away from rendering

So yeah, sounds like you are the one full of bovine feces here
 

SlimySnake

Flashless at the Golden Globes
Not having looked into it, that sounds like a Watch Dogs problem? There are other games that they perform better than their expected comparison, and many at about what you'd expect, depends on if the developer tailored for Nvidia or AMD hardware more. On ray tracing, yes, they definitely perform a tier down.

Will it not be as good as the tensor core approach, look, I could definitely believe it. But I can also see there being an opportunity for compute to asynchronously boost the perceived render resolution without dedicated cores for it.
ray tracing performance in ALL RT games is trash compared to 30 series or even 20 series cards.


This is their 3080 competitor performing like a 2080 super.
 

assurdum

Banned
The PS5 uses a custom AMD chip running on a custom OS. Maybe what the Fidelity FX uses for the machine learning is in use for something else in the system and cannot be used for up-scaling? Just a thought
AMD didn't said won't uses machine learning at all? (though Cerny already stated machine learning is available inside the ps5 GPU in Road to the ps5).
 
Last edited:

DeaconOfTheDank

Gold Member
Bovine feces.
There is no "tensor core dedicated to DLSS".
Nor is there anything magical about "tensor cores" themselves, essentially, it is "a bunch of FP ops".

And the operation used to improve what otherwise is 90%+ TAA is called "neural network inference". GPUs are pretty darn good at it out of the box.
Tensor cores are actually hugely instrumental to ML-related tasks. In particular, they provide hardware acceleration for matrix multiplication operations - this is the very foundation of ML. This is where the "tensor" keyword comes from. With compute cores, they are primarily used for parallel graphics rendering tasks. They can be programmed to perform matrix multiplication required by ML, but the general purpose nature of a compute core is not optimized for this particular task.

Source: My job has me building, training, and deploying ML models on a regular basis.
 
Last edited:

regawdless

Banned
Tensor cores are actually hugely instrumental to ML-related tasks. In particular, they provide hardware acceleration for matrix multiplication operations - this is the very foundation of ML. This is where the "tensor" keyword comes from. With compute cores, they are primarily used for parallel graphics rendering tasks. They can be programmed to perform matrix multiplication required by ML, but the general purpose nature of a compute core is not optimized for this particular task.

Source: My job has me working with building, training, and deploying ML models on a daily basis.

A. Tensor cores are much smaller and faster than CUDA cores at the cost of reduced precision. 119 Tensor FP16 TFLOPS for the 3080. Up to double that for sparse matrices. So that is a pretty huge performance boost for workloads that are amenable to it (e.g. inferencing)

B. Also has the benefit of using a separate, dedicated piece of silicon rather than using the standard CUDA cores/Stream Processors and thus taking some performance away from rendering

So yeah, sounds like you are the one full of bovine feces here

Thanks for the good clarifications. Won't stop him from spreading his nonsense though, unfortunately.
 

DonkeyPunchJr

World’s Biggest Weeb
Thanks for the good clarifications. Won't stop him from spreading his nonsense though, unfortunately.
Just sour grapes from the usual suspects. Any time there’s a post about DLSS or ray tracing there are the same 3 folks who make it their life’s mission to convince people they shouldn’t want it.

You just know if the positions were reversed and these were AMD exclusive features, these same fanboys would be praising it to the heavens.
 

Excess

Member
Just sour grapes from the usual suspects. Any time there’s a post about DLSS or ray tracing there are the same 3 folks who make it their life’s mission to convince people they shouldn’t want it.

You just know if the positions were reversed and these were AMD exclusive features, these same fanboys would be praising it to the heavens.
Or the very simple fact that Moore's Law is dead, and that comparable hardware by itself won't save AMD from competing with Nvidia in the GPU space. AMD absolutely needs an answer to DLSS.
 
I want to hold out hope that AMD will have something that will come within swinging distance to DLSS, but I am not too sure. RTX is cool, but DLSS has been a game changer and probably one of the more fascinating techniques in GPU trickery in recent memory. I'd settle for half of what DLSS could offer on consoles, would be cooler if they would standardize Super Resolution/DLSS into one acronym and make it easier for game developers.
DLSS was not good at launch. It took Nvidia months of testing based on feedback to make it good. I expect same from MS DirectX ML and AMD super resolution.
 

llien

Member
AMD will always be behind Nvidia, although they did beat the crap outta Intel old ass cpus that pulls 250w with 8cores
You have written this post in times when NV was forced to use GA102 as 3080 and end up in ridiculous situation of 3060 having more VRAM than 3060Ti, 3070 and 3080.

AMD is "always behind" only in green bubbles.
SAM and anti-lag are two bright examples from this very generation, on top of RDNA2 kicking Ampere's arse.
 

llien

Member
Tensor cores are much smaller and faster than CUDA cores at the cost of reduced precision.
Bovine feces.
Tensor cores are not faster, let alone "much faster" than normal shaders (which you referred to as "CUDA cores").
It's not about fp precision either.

It's the fact that shadres are more like generic compute units, sort of small and (compared to x86 world) dumb CPUs.
Whereas tensor cores are like AVX support in CPUs. They can only do that bunch of fp ops. (4 x 4 "matrix", but in reality, you could view it as simply 16 ops)
So, normal shader can do many things, while tensor core can do only that one thing.
In the past shaders could do int + fp op per cycle.
With Ampere, NV shaders can do int + fp or fp + fp.
That is why Huang thought it is a wise idea to claim that ampere has twice the number of shaders it has.
Oh well, of all the bazinga marketing coming out of it, this is perhaps the least harmful... :messenger_relieved:

Also has the benefit of using a separate, dedicated piece of silicon rather than using the standard CUDA cores/Stream Processors and thus taking some performance away from rendering
I see, "separate" is how things get "better". Will remember that, thanks for the insight. :messenger_beaming:

And to math, 3060 has 120 "tensor cores" and 3584/2 => 1792 shaders.
So, shaders could do 2*1792 while "tensor cores" can do 16*120 => roughly half the fp power of shaders.
Which, indeed, isn't insignificant, although, it's not clear if tensor cores are indeed independent or share some part of ALU units.with normal shaders.

On top of it, low end GPUs with tensor cores are still overpowered by higher end GPUs from earlier gens, so the "uh, you need these to apply some NN post processing to the TAA(u) upscaling also known as DLSS 2.0" is... likely a lie.
 
Last edited:

Krizalidx11

Banned
You have written this post in times when NV was forced to use GA102 as 3080 and end up in ridiculous situation of 3060 having more VRAM than 3060Ti, 3070 and 3080.

AMD is "always behind" only in green bubbles.
SAM and anti-lag are two bright examples from this very generation, on top of RDNA2 kicking Ampere's arse.

When did that happen? no cards to be found and it couldn't beat ampere in benchmarks !!!!


not to mention the crappy control panel that won't solve audio recording issues if i'm using a 3.5mm jack headphones
 

llien

Member
When did that happen?
Q4 last year, when Huang was forced to use GA102 as 3080 chip.

no cards to be found
Actually, mindfactory sales show that there is quite a stock of GPUs, prices are way about MSRP though.

it couldn't beat ampere in benchmarks !!!!
Top cards (e.g. 6900XT vs 3090) traded blows over wide range of benchmarks, AMD was notably better positioned with the newer games.
But that doesn't matter and I guess you've missed the point.
 

ratburger

Member
4A Games, the developers of Metro Exodus, don't seem to believe it's very good.

Metro-Exodus-FAQ-AMD-FSR.png


FAQ: https://www.metrothegame.com/news/the-metro-exodus-pc-enhanced-edition-arrives-may-6th

Likely going to be an additional year or more before AMD can hit DLSS 2.0 level of quality.
 
Last edited:

LordOfChaos

Member
4A Games, the developers of Metro Exodus, don't seem to believe it's very good.

Metro-Exodus-FAQ-AMD-FSR.png


FAQ: https://www.metrothegame.com/news/the-metro-exodus-pc-enhanced-edition-arrives-may-6th

Likely going to be an additional year or more before AMD can hit DLSS 2.0 level of quality.



We have been contacted by Deep Silver/4A Games representative with the following statement:

4A Games has not evaluated the AMD FidelityFX Super Resolution feature for Metro Exodus at this time. In our FAQ, we were referring to the AMD FidelityFX open source image quality toolkit which targets traditional rendering techniques that our new RT only render does not use, and noting that we have our own Temporal Reconstruction tech implemented natively which provides great quality benefits for all hardware, so do not currently plan to utilize any other toolkits.
4A Games is always motivated to innovate, evaluate, and use the newest technologies that will benefit our fans across all platforms and hardware.

We knew FidelityFX wasn't the same already. They have not evaluated the new Super Resolution, a new feature being added, that they have not tested
 
Last edited:

VFXVeteran

Banned
We have been contacted by Deep Silver/4A Games representative with the following statement:



We knew FidelityFX wasn't the same already. They have not evaluated the new Super Resolution, a new feature being added, that they have not tested
They more than likely have the API for it. But 4A doesn't need it. It's a software solution anyway and they know their graphics engine better than anyone.
 

LiquidMetal14

hide your water-based mammals
I'm interested to see how their 1.0 will be vs what nvidia has because nvidia has spent millions in R&D and a long while to get to the 2.0 implementation we have now on PC.

Squeezing that extra performance at not cost is always welcome.
 

truth411

Member
What make you think it will be as good as DLSS 2.0?

Even if its not and say with AMD verison you have to have a little more resolution to equal DLSS 2.0 performance, thats still a huge performance savings.
My Guesstimate

AMD FX 1600p = Nvidia DLSS 1440p In performance.

Thats still a huge performance savings compared to native 4k. And the performance savings can be repurposed elsewhere.
 

Rikkori

Member
There is no reason why it can't be the same qualitatively as DLSS 2.0. The question is more about the performance difference and whether they'd take the performance or the quality. The only thing DLSS does anyway is the clamping step, which is what the ML is trained for. That's it. Then there's a cost to run that model on the GPU during runtime, which is where the performance difference would emerge (due to no tensor cores). Though even here it's not necessarily a given that FSR would be slower, after all RDNA 2 has the edge over Ampere in FP16 (3080 = 30 TF vs 6800 XT = 42 TF FP16).

That being said, if it's the same quality but let's say it lacks 10% of the performance saving of DLSS, is anyone gonna cry over that? Naaah.
 
I doubt it will have much impact.

There are already various reconstruction methods available which produce a reasonably convincing 4k image from a roughly 50% base.

And nobody thinks it's gonna be as good as DLSS, so it's unlikely to be SIGNIFICANTLY better than something like checkerboarding.

So I'm not sure there's really a need for another offering.

At best it'll just mean slightly more convincing reconstructions (ie no added performance) and an off the shelf solution for studios still just using 1440p.

I hope I'm wrong but I doubt it's gonna be game changing.
well DLSS is undoubtably game changing so anything in that same ballpark will be a massive boon to consoles
 

Hunnybun

Member
well DLSS is undoubtably game changing so anything in that same ballpark will be a massive boon to consoles

Well from what I understand DLSS offers 4x the base resolution for results significantly better than checkerboarding (2x the base res).

I guess if AMD's thing could get something like checkerboarding at maybe 3x the base res then that would be a worthwhile improvement. Guess we'll see.
 
Top Bottom