Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[Dezii90]

It's actually funny how people are grateful that someone (game devs and nvidia) cut their legs fist with a TAA chainsaw, and then offered (for a price) a high-tech prostesis that does almost the same, but is inherently twitchy (TAA jitter) and does not follow the commands immediately (frame accumulation). All sorts of temporal upscaling are the worst thing that happened to the gaming industry (well, maybe it's for the best, there are other hobbies too or older games)

TAA was the best option to handle new post processing for a lot of engines thats why its became default. No one wants to really experiment with new aa to replace taa. Fxaa + smaa should at least be offered most of the time.

 

[Jan Olšan]

Upscaling will never be better than native. The only thing that allows them to claim that crap is various extra filtering like the temporal reconstruction... which, guess what, could/should be applied on native resolution if it helps, and is not inherent part of the upscaling much like their "reflex" is not an inherent part of frame generation (something they love to imply in their denial arguments about latency).

If the lipstick is any good, put it on a girl and not just the pig, and then dare to compare them.

 

[basix]

Well, what is native? No AA? MSAA? TAA? FXAA? SMAA? DLAA? FSR native? Considering that many engines render at half or quarter rate for some subsystems and with halfed framerate on top, it is a tricky question.

It is true, that TAA introduced many problems but it solved some others. But there was never any AA technique in history, which was perfect. Only high degrees of Super Sampling come close to perfection.
And it is also true, that if you increase resolution with a given AA technique (not changing it) will yield in better image quality. And what does Super Sampling do? Exactly that, increase render resolution. DLAA will always look better than DLSS-P.
The biggest TAA flaws are, that it gets blurry in motion and even more blurry if you reduce the render/output resolution. And I do not like blurry images at all.

But here is the key point of DLSS and Co:
It is actually supersampling. If done right, it can look better than "native". And approaching true supersampling quality. DLSS3-Q did look better than "native TAA" in many games, because normal TAA has the very same flaws as any other temporal AA technique (ghosting, smearing, blurring in motion). Was it perfect? By far no. Even at 4K output I still used DSR/DLDSR because of the image blurring. But in many cases it was better what was available as alternative (often only TAA).

And that is why DLSS4 is a game changer for me. It resolves 90% of the image blurring stuff. The difference to DLSS3 is night and day. DSR/DLSR not needed anymore. Difference between DLSS-P and DLAA is rather marginal regarding image sharpness. I mean, the new DLSS4-P often beats the old DLSS3-DLAA in many aspects of the image. Sure, DLAA vs. P still resolves fine detail better etc. but that is to be expected. And at the same time DLSS4 improves general AA quality and most other TAA quirks (ghosting and smearing) are vastly reduced as well. So DLSS4 is a huge upgrade and should be praised by all gamers: We have finally overcome the dark ages of TAA blur!

HWUB made a great video about it:

There were also comparisons somewhere of DLSS4 vs. SSAA and DLSS4 clearly won. DLSS with its huge history buffer (FSR3 maxes out at 72 frames I think) just has more information to to work with than if you use normal 2x2 Supersampling.

So my big hope is, that AMDs FSR4 gets close to DLSS4 regarding quality. And that it is available on as many GPU architectures as possible.

 

[poke01]

So my big hope is, that AMDs FSR4 gets close to DLSS4 regarding quality. And that it is available on as many GPU architectures as possible.

My hope too, DLSS is the major psychological barrier that needs to be broken. Because people use that as excuse to say “oh it has less VRAM but you get DLSS.”

Nope not anymore, in the future it should be “FSR4 etc is just as good and if you go AMD it’s open source and you get more VRAM.”

 

[basix]

Yes, that is another aspect. I hope AMD knows as well, that FSR4 is key in gaining mindshare and market share.

 

[adroc_thurston]

My hope too, DLSS is the major psychological barrier that needs to be broken. Because people use that as excuse to say “oh it has less VRAM but you get DLSS.”

Nope not anymore, in the future it should be “FSR4 etc is just as good and if you go AMD it’s open source and you get more VRAM.”

Irrelevant, the shilling will move onto something else.
What AMD needs is a boomstick.

 

[Tup3x]

Latest iteration of DLAA pretty much fixes TAA issues and I really hope that native FSR4 can do the same (upscaling versions too).

 

[exquisitechar]

Yes, that is another aspect. I hope AMD knows as well, that FSR4 is key in gaining mindshare and market share.

Doubt FSR4 will get close to DLSS4. It’ll probably catch up to DLSS3 at best and most people will then say that they’re buying Nvidia because DLSS4 is better, among many other must-have Nvidia exclusive features.

I think DLSS4 is great, but I doubt that Radeon’s sales would improve much even if FSR4 matched it. Adroc is right, AMD needs to smash Nvidia with a halo part + consistently release competitive generations for a while to shake up the market. If it’s not DLSS4 then it will be CUDA, the encoder or whatever the new shiny thing is, because people decide to buy Nvidia by default and justify why afterwards.

 

[DAPUNISHER]

Irrelevant, the shilling will move onto something else.
What AMD needs is a boomstick.

Nvidia shooting itself in the foot with their own boomstick, helps too.

Latest iteration of DLAA pretty much fixes TAA issues and I really hope that native FSR4 can do the same (upscaling versions too).

If AMD is on their usual pace, they are a generation behind.

 

[adroc_thurston]

Nvidia shooting itself in the foot with their own boomstick, helps too.

No.
AMD just has to do what they do best (3D stacked voodoo).

AMD is on their usual pace, they are a generation behind

It's not a "generation" thing, it's "how slow do you want your upscaler to be".

 

[poke01]

AMD just has to do what they do best (3D stacked voodoo).

Does NV suck at this SoIC stuff?

 

[Gideon]

AMD just has to do what they do best (3D stacked voodoo).

All for it, as long as it isn't the Voodoo of 3dfx kind ...

 

[adroc_thurston]

Does NV suck at this SoIC stuff?

Everyone not AMD does.

All for it, as long as it isn't the Voodoo of 3dfx kind ...

Lmao yeah

 

[basix]

Doubt FSR4 will get close to DLSS4. It’ll probably catch up to DLSS3 at best and most people will then say that they’re buying Nvidia because DLSS4 is better, among many other must-have Nvidia exclusive features.

I think FSR4 Preview (aka "research project" at CES) looked better than DLSS3. But we will see in a few days, how things will turn out.

But if you listen to most gamers, they care much more about Perf/$, good framerates and decent upscaling than FG, MFG and maxed out raytracing.

I mean, you can compare a 650$ card with slightly worse FSR4 vs. a 1000$ card (if you can find one now at that price) with slightly better DLSS4. Perf/$ of the first card still wins by far.

But if the herd is stupid (or chooses to be stupid) you can't win anyways. You cannot win arguments against dumb people ;-)

 

[adroc_thurston]

But if you listen to most gamers, they care much more about Perf/$, good framerates and decent upscaling than FG, MFG and maxed out raytracing.

They cared about NVENC before.
The goalposts will move towards whatever thing AMD doesn't have/is worse at.

 

[tsamolotoff]

TAA was the best option to handle new post processing for a lot of engines

It's not the best, it's just the one that allows devs to expend the least amount of resources at the expense of my eyes. You can just fire up some old game like free HL2: Lost coast demo and compare how clear the image is as compared to something like the upcoming Talos Principle 1: UE5 smearslop edition (also free demo available). I don't care how many light sources TAA vaseline allows to display, if it's all blurry mess I simply won't play that. No amount of filtering and AI can bring back lost details and remove jitter (which makes whole picture unstable)

But there was never any AA technique in history, which was perfect.

It's not a reason to saddle everyone with temporal solution while providing no other options.

There were also comparisons somewhere of DLSS4 vs. SSAA and DLSS4 clearly won.

Is there a game that supports both?

 

[tsamolotoff]

Upscaling will never be better than native.

The problem is that there no native anymore - as TAA and its derivatives captured the whole market, now it is used to hide the fact that the current technologies can't really handle whatever technical (not graphical, mind you) feats the devs and IHVs want to push (like RTRT and whatnot).

As the result, the progress in graphics, as Alex Sannikov from GGG put it, basically halted to a stop since RTX was introduced. Instead of using 'good enough' and clever tricks like prebaked lightning (which is also RT, just not real-time one) that allow you to get the same (or even better) picture, the devs are now trying to generate the picture in a bruteforce way with universal, but very inefficient methods. It'd be fine if we as consumers didn't have to pay for that with our wallets and eyes.

 

[poke01]

357mm2 according to videocardz

 

[inquiss]

I think FSR4 Preview (aka "research project" at CES) looked better than DLSS3. But we will see in a few days, how things will turn out.

But if you listen to most gamers, they care much more about Perf/$, good framerates and decent upscaling than FG, MFG and maxed out raytracing.

I mean, you can compare a 650$ card with slightly worse FSR4 vs. a 1000$ card (if you can find one now at that price) with slightly better DLSS4. Perf/$ of the first card still wins by far.

But if the herd is stupid (or chooses to be stupid) you can't win anyways. You cannot win arguments against dumb people ;-)

The psychology is generally not "what is the best value for me based on the inputs" it's "I'm going to buy Nvidia how do I justify my purchase". This is also, more or less, how psychology works for many decisions. We use post rationalization to trick our brains into thinking we made the right decisions on the first place. It works the opposite way around to how many people think it does.

 

[Timorous]

They cared about NVENC before.
The goalposts will move towards whatever thing AMD doesn't have/is worse at.

I suspect it will move to the fact AMD hardware actually has the correct spec and that will be an issue some how because the RMA process is a can't miss out on NV feature.

 

[poke01]

They cared about NVENC before.
The goalposts will move towards whatever thing AMD doesn't have/is worse at.

Why did they care is the question? Was NVENC better than AMDs encoders in terms of speed and quality?

 

[poke01]

People care about encoders on GPUs, that’s why AMD improved them for RDNA4

 

[Thunder 57]

Why did they care is the question? Was NVENC better than AMDs encoders in terms of speed and quality?

AMD'x x264 sucks. x265/AV1 is fine though. I prefer CPU encoding regardless. Not exactly a streamer.

 

[inquiss]

Why did they care is the question? Was NVENC better than AMDs encoders in terms of speed and quality?

They cared because Nvidia influenced them to care. No other reason.

 

[basix]

Is there a game that supports both?

Not sure what game it was. Maybe it was just 8K downscaled to 4K vs. 4K with DLSS.
OGSSAA is worse than SGSSAA but latter has extinct a long time ago.

I think the "newer" Hitman games can use DLSS and have also an in-built supersampling slider (100....200%)

The psychology is generally not "what is the best value for me based on the inputs" it's "I'm going to buy Nvidia how do I justify my purchase". This is also, more or less, how psychology works for many decisions. We use post rationalization to trick our brains into thinking we made the right decisions on the first place. It works the opposite way around to how many people think it does.

Completely agree. Many people seek for justifications. It's somewhat natural, you do not want to feel bad if you spend money on something. But in some cases, the reality just gets bent a little bit (too) far and you begin lying to yourself and in the end hurting yourself. E.g. spend too much money on something you don't need.