• 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.
  • The Politics forum has been nuked. Please do not bring political discussion to the rest of the site, or you will be removed. Thanks.

AMD’s Equivalent to DLSS, FidelityFX Super Resolution Is Coming to PS5 & Xbox Series Consoles

RockOn

Member
Feb 18, 2021
108
129
250
UK
Well Cerny only said they have the same support for accelerated RT HW. He never once said they accelerated the HW for ML. That’s was Microsoft who made that claim. The algorithms Sony can use for Super Resolution can be trained to run at FP16 effectively with API from AMD. Devs will be able to use the same ones across the board to avoid having to use different methods and that saves money and time. If they want to use DirectML from Xbox they will need to use that API. That will have the algorithms trained by the work with Azures and Nvidias collaboration. You can look into the AI work they did and it makes sense why Xbox would want to combine the tech into its next gen systems. AMD SR aims to be a one stop for all solution for devs. Imagine they use the AMD version across all consoles for a predictable effect within a budget.
AMD have OpenML which is an open source project. So will work on any platform
 
Last edited:

RockOn

Member
Feb 18, 2021
108
129
250
UK
RDNA1 has it in the Whitepaper for the architecture, some of the CUs have the ability to scale below FP16 in Integer Size, down to FP8 and FP4.


"Some variants of the dual compute unit expose additional mixed-precision dot-product modes in the ALUs, primarily for accelerating machine learning inference. A mixed-precision FMA dot2 will compute two half-precision multiplications and then add the results to a single-precision accumulator. For even greater throughput, some ALUs will support 8-bit integer dot4 operations and 4-bit dot8 operations, all of which use 32-bit accumulators to avoid any overflows."
Mixed integer ia a standard feature of RDNA 2 enhanced CUs
 

RockOn

Member
Feb 18, 2021
108
129
250
UK
Well Cerny only said they have the same support for accelerated RT HW. He never once said they accelerated the HW for ML. That’s was Microsoft who made that claim. The algorithms Sony can use for Super Resolution can be trained to run at FP16 effectively with API from AMD. Devs will be able to use the same ones across the board to avoid having to use different methods and that saves money and time. If they want to use DirectML from Xbox they will need to use that API. That will have the algorithms trained by the work with Azures and Nvidias collaboration. You can look into the AI work they did and it makes sense why Xbox would want to combine the tech into its next gen systems. AMD SR aims to be a one stop for all solution for devs. Imagine they use the AMD version across all consoles for a predictable effect within a budget.
They don't need DirectML, Machine Learning are libraries and can be used on any Platform. AMD has OpenML. Your making more of DirectML. Not every platform need that, ther are other API for ML/AI etc
 

MrSec84

Member
Jul 1, 2020
111
150
230
Mixed integer ia a standard feature of RDNA 2 enhanced CUs
My point was that going below FP16 has been possible on a portion of even RDNA1 ALUs, so this whole Machine Learning feature on XBox Series hardware isn't unique to it, it's baked into RDNA1.
From a software perspective you could still pack multiple FP8 tasks into a single FP32 ALU, it would just need to be a integrated feature in your game engine as a developer.

What I'm not sure of is whether RDNA2 allows for all of it's ALUs to handle any variation down to FP4 or whether like RDNA1 it's only baked into a fraction of the whole GPU's ALUs, so like a quarter of the ALUs in the system can do it, it would make sense for all ALUs to be able to handle it so that all code could run this way of needed, to maximise GPU performance.
 
  • Like
Reactions: RockOn

Omni_Manhatten

Neo Member
Feb 22, 2021
36
40
100
Algorithms take hundreds of thousands of man hours to train. If you think DLSS is some easy tech that just requires a run of the mill algorithm I highly encourage you to do research on the topic. The work done with Nvidia and Azure was indeed vast and it will not be a flick of the switch easy to apply algorithm that just goes head to head with DLSS. Yes it is very impressive what MS done.
 

RockOn

Member
Feb 18, 2021
108
129
250
UK
Algorithms take hundreds of thousands of man hours to train. If you think DLSS is some easy tech that just requires a run of the mill algorithm I highly encourage you to do research on the topic. The work done with Nvidia and Azure was indeed vast and it will not be a flick of the switch easy to apply algorithm that just goes head to head with DLSS. Yes it is very impressive what MS done.
Never said it didn't. But you making out its only MS using Machine Learning and AI. AMD/sony will be using it too
 
  • Like
Reactions: polybius80

RockOn

Member
Feb 18, 2021
108
129
250
UK
My point was that going below FP16 has been possible on a portion of even RDNA1 ALUs, so this whole Machine Learning feature on XBox Series hardware isn't unique to it, it's baked into RDNA1.
From a software perspective you could still pack multiple FP8 tasks into a single FP32 ALU, it would just need to be a integrated feature in your game engine as a developer.

What I'm not sure of is whether RDNA2 allows for all of it's ALUs to handle any variation down to FP4 or whether like RDNA1 it's only baked into a fraction of the whole GPU's ALUs, so like a quarter of the ALUs in the system can do it, it would make sense for all ALUs to be able to handle it so that all code could run this way of needed, to maximise GPU performance.
I agree, and dot 4/dot 8 is also supported. Just newer RDNA 2 have the hardware acellerators
 

Omni_Manhatten

Neo Member
Feb 22, 2021
36
40
100
Never said it didn't. But you making out its only MS using Machine Learning and AI. AMD/sony will be using it
I never said that. You’re ( you are) arguing with yourself. I’m telling you it takes years to train an algorithm correctly. It’s not just some but AMD can just throw a comparable algorithm to DLSS on PS5 and get that result. Especially at FP16 which is what Super Resolution will be using on Sony’s machine. As Andrew Goosen said they had to invest to support ML on a HE level. This was reiterated by Jason Ronald just yesterday on the Lords podcast and he even mentioned MS has done work on Super Resolution beyond that of DLSS. Which makes sense because MS and Nvidia have been doing that work for years. Which is what I said to you. It is in fact impressive and not some run of the mill algorithm.
 

Omni_Manhatten

Neo Member
Feb 22, 2021
36
40
100
I agree, and dot 4/dot 8 is also supported. Just newer RDNA 2 have the hardware acellerators
The RDNA2 papers show gaming support standard is FP16. Period. Integer level support must be added.
I agree, and dot 4/dot 8 is also supported. Just newer RDNA 2 have the hardware acellerators
You need to go back and check again. It’s pretty clearly outlined what int support for what is in RDNA 2. I mean it’s a very good break down. Nvidia uses Tensor cores to do the ML work and RDNA has no equivalent to this. MS and AMD worked together to allow Int 8/4 support. They had to work for the cores to support calculations at that level. This was specifically to mimic tensor cores for ML work.

 

MrSec84

Member
Jul 1, 2020
111
150
230
I agree, and dot 4/dot 8 is also supported. Just newer RDNA 2 have the hardware acellerators
There's no separate acceleration technology for ML in RDNA Architecture, RDNA1 has some of it's ALUs able to reduce Integer Sizes to FP4 and 8, I believe RDNA2 can do it across all ALUs, things like improve cache, more efficient overall design, but no specific processing hardware dedicated to Accelerating AI Code is present.

Super Resolution will be handled on lower integer pathways in the GPU, it will use the RDNA2 ALUs, alongside Graphics, Physics, Audio, etc processing.
 

RockOn

Member
Feb 18, 2021
108
129
250
UK
The RDNA2 papers show gaming support standard is FP16. Period. Integer level support must be added.

You need to go back and check again. It’s pretty clearly outlined what int support for what is in RDNA 2. I mean it’s a very good break down. Nvidia uses Tensor cores to do the ML work and RDNA has no equivalent to this. MS and AMD worked together to allow Int 8/4 support. They had to work for the cores to support calculations at that level. This was specifically to mimic tensor cores for ML work.


 

//DEVIL//

Member
May 28, 2014
2,420
1,871
690
Their approach is not a toggle switch in the AMD software as far as what I understood . It has to be game per game bases . If that’s the case, it’s DOA day zero.

Nvidia market more games than AMD. Specially the big games too. Add to that dlss is unreal engine toggle . So yeah ...
 
Last edited:

Zathalus

Member
Jul 16, 2020
1,094
2,841
470
Their approach is not a toggle switch in the AMD software as far as what I understood . It has to be game per game bases . If that’s the case, it’s DOA day zero.

Nvidia market more games than AMD. Specially the big games too. Add to that dlss is not unreal engine toggle . So yeah ...
If the AMD version is supported by both consoles then the Nvidia solution will inevitably end up getting less support in the long run.
 

vkbest

Member
Jan 23, 2017
1,729
1,687
500
When Digital Foundry says it's checkerboard 4K on PS4 Pro, then I take them at their word.

Even if DF said, its not checkerboard, they are rendering at 1920x2160 and stretching the image to 4k. That is a scaling technique, Checkerboard is a reconstruction technique. That is because is so blurry.
 
Last edited:

sncvsrtoip

Member
Jul 18, 2019
739
1,205
405
With big probabilities ps5 doesn't support int8/4 with higher performance (David Cage about xsx advantage in ml when only thing ms said connected to ml is int8/4) so it will just use fp16/int16 calculations with their version of amd super resolution with worse performance but we don't know yet what performance cost of it is (maybe neglectable) or quality of amd soultion (to be honest I would rather expect dlss1.0 quality and not dlss2.0)
 
Last edited:
  • Thoughtful
Reactions: ErRor88

RockOn

Member
Feb 18, 2021
108
129
250
UK
With big probabilities ps5 doesn't support int8/4 with higher performance (David Cage about xsx advantage in ml when only thing ms said connected to ml is int8/4) so it will just use fp16/int16 calculations with their version of amd super resolution with worse performance but we don't know yet what performance cost of it is (maybe neglectable) or quality of amd soultion (to be honest I would rather expect dlss1.0 quality and not dlss2.0)
PS5 isnt RDNA1, its RDNA2. as you can see RDNA2 from the table above supports mixed integer int4/int8 & dot4/dot8. Why can't people read and stop spreading the fud.

2nd PS5 Wired Article mentions 'Machine Learning'
QUOTE
"I could be really specific and talk about experimenting with ambient occlusion techniques, or the examination of ray-traced shadows," says Laura Miele, chief studio officer for EA. "More generally, we’re seeing the GPU be able to power machine learning for all sorts of really interesting advancements in the gameplay and other tools." Above all, Miele adds, it's the speed of everything that will define the next crop of consoles.
 

sncvsrtoip

Member
Jul 18, 2019
739
1,205
405
PS5 isnt RDNA1, its RDNA2. as you can see RDNA2 from the table above supports mixed integer int4/int8 & dot4/dot8. Why can't people read and stop spreading the fud.

2nd PS5 Wired Article mentions 'Machine Learning'
QUOTE
"I could be really specific and talk about experimenting with ambient occlusion techniques, or the examination of ray-traced shadows," says Laura Miele, chief studio officer for EA. "More generally, we’re seeing the GPU be able to power machine learning for all sorts of really interesting advancements in the gameplay and other tools." Above all, Miele adds, it's the speed of everything that will define the next crop of consoles.
its hybrid (doesn't have hw vrs for example), this EA ml talks is interesting (also similar comments from Sega) but imho there is lack of int8/4 support
 

RockOn

Member
Feb 18, 2021
108
129
250
UK
its hybrid (doesn't have hw vrs for example), this EA ml talks is interesting (also similar comments from Sega) but imho there is lack of int8/4 support
Nonsense, ther is hardware VRS/Ray Tracing & mixed integer int/4/in8 support all them features are basic RDNA 2 features(and will be in PS5) PS5 won't support tier2(as that MS software)
More talk on PS5 doing Machine Learning
 
Last edited:

sncvsrtoip

Member
Jul 18, 2019
739
1,205
405
Nonsense, ther is hardware VRS/Ray Tracing & mixed integer int/4/in8 support all them features are basic RDNA 2 features(and will be in PS5) PS5 won't support tier2(as that MS software)
More talk on PS5 doing Machine Learning
nah, but will not waste my time on it, belive what you want, we have already enough knowledge (multiple devs talking, soc die photos) thats we know its some kind of hybrid
 
  • Like
Reactions: ErRor88

small_law

Member
Nov 30, 2017
410
370
420
I've been looking for a place to leave a confession, more of an apology.

I was planning on getting an 11th gen core i9. Went to micro center to look at z590 motherboards and there sat a 5950X. I thought why wait and got it. I just finished putting together my first AMD build.

You team red guys were right. Intel's two generations behind on desktop processors. It's not close. It makes the 9900K in my older build look twice as old as it is.

I won't concede graphics cards, but there's only one choice for desktop CPU right now.
 
  • Like
Reactions: marquimvfs

jonnyp

Member
Jan 12, 2008
3,386
2,878
1,290
Nobody knows how good it is or isn't. Could be better then DLSS 3.0 or even 4.0, could be absolute dog shit.

Its funny they choose linus for this reveal, guess shilling for AMD worked out well for him.

He has 13 mill subscribers, that's why.
 

mitchman

Member
Aug 20, 2013
2,044
1,513
720
Oslo, Norway
I've been looking for a place to leave a confession, more of an apology.

I was planning on getting an 11th gen core i9. Went to micro center to look at z590 motherboards and there sat a 5950X. I thought why wait and got it. I just finished putting together my first AMD build.

You team red guys were right. Intel's two generations behind on desktop processors. It's not close. It makes the 9900K in my older build look twice as old as it is.

I won't concede graphics cards, but there's only one choice for desktop CPU right now.
Yepp. I'm 100% Intel CPU-free on both desktop and my laptop now. 5950x on desktop, Ryzen 9 4900HS on my laptop. First AMD on both since the Athlon 64 days. Congrats!
 
Last edited:
  • Like
Reactions: small_law