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

 

[SolidQ]

To break the cycle, they either need a better product

if RDNA4 will have between 7900XT/XTX and 4080RT, also AI upscaler for 500-550$ + good marketing, that's enought to get market share

 

[Mahboi]

Not happening. We were talking of raster somewhere above 7900XT and lower than XTX.
I also made a napkin math expectation based on that raster somewhere and found that it should realistically be around a 4070 Ti.
Even if you don't expect much out of Blackwell, it will be a walk in the park for Nvidia to offer better. Heck the 4090 already trounces it, put it on N4X and give a 4090-ish card with 20% less power draw, that's it, they won.

 

[igor_kavinski]

if RDNA4 will have between 7900XT/XTX and 4080RT, also AI upscaler for 500-550$ + good marketing, that's enought to get market share

Nope. Don't see that happening.

 

[marees]

I also made a napkin math expectation based on that raster somewhere and found that it should realistically be around a 4070 Ti.

My expectationis that navi 44 (equivalent in chip size to 6500 xt / polaris 14) should match the 4060 ti 8gb in both raster & RT, hence probably $300

The navi 48 chip is exactly double the navi 44 /4060 ti 8gb. Where does that land 4070 ti or 4070 ti super ? Atleast overclocked board partner versions should match the 4070 ti super

In terms of competition nvidia could just rebrand the 4060 ti as 5060.

 

[SolidQ]

hence probably $300

DOA for 8GB

Don't see that happening.

isn't small die size make to cheaper it? Unless AMD is greedy

 

[igor_kavinski]

Unless AMD is greedy

This. Intel needs to step up their game to make them less greedy. They know people who can't afford Nvidia have no choice but to turn to their GPUs, unless they want to play lottery with Intel ARC GPUs (will the game run flawlessly??? Will it or won't it??? I hope it does. It will be so sad if it doesn't!).

 

[SolidQ]

Intel needs to step up their game to make them less greedy

Intel GPU for now is a pig in a poke, but we will to see in 4-5 years.
AI upscaler aka DLSS is now main reason people choose NVIDIA. Games now tends to non-removable upscaler. That what AMD need

 

[Mahboi]

DOA for 8GB

Not really, 1080p games requiring over 8Go of textures is still basically none of them.
Gets more hairy when you reach 2K.

isn't small die size make to cheaper it? Unless AMD is greedy

AMD is greedy.

 

[SolidQ]

1080p games requiring over 8Go of textures is still basically none of them.

ok for now, but what about 2025-2026 etc

 

[poke01]

AMD needs better ray tracing in all types, better upscaler and be more efficient AND be faster than Nvidia and cheaper, only then will gamers jump ship.

 

[poke01]

ok for now, but what about 2025-2026 etc

The latest games Wukong is not VRAM hungry, UE5 has tools that make it so.

 

[igor_kavinski]

AI upscaler aka DLSS is now main reason people choose NVIDIA.

XeSS is already superior in many ways to AMD FSR. Intel just needs to put slightly more effort and they could win significant mindshare (once online forums everywhere agree that FSR is inferior to both DLSS and XeSS, it will either force AMD to improve FSR or reduce their card prices).

 

[marees]

it will either force AMD to improve FSR

Hopefully AMD is working on AI enabled FSR that can launch alongside RDNA 4

 

[poke01]

Hopefully AMD is working on AI enabled FSR that can launch alongside RDNA 4

They are not. Why? If AMD did then Sony wouldn't work on their own upscaler for PS5 Pro. RDNA5? Very likely we will see AI upscaling in that gen.

 

[soresu]

Hopefully AMD is working on AI enabled FSR that can launch alongside RDNA 4

They might paper launch it with RDNA4 only to hard launch it months later as they did with FSR3.

 

[soresu]

Why? If AMD did then Sony wouldn't work on their own upscaler for PS5 Pro

Sony want to be certain that the quality of the upscaler meets their standards whatever the meat of the scaling system is.

Said upscaler will probably be integrated with the Playstation SDK to minimise effort implementing it on the game dev side of things.

 

[igor_kavinski]

Sony want to be certain that the quality of the upscaler meets their standards

Seems like their data shows not many gamers using the 60fps performance option due to the awful shimmering effects.

 

[inquiss]

Intel GPU for now is a pig in a poke, but we will to see in 4-5 years.
AI upscaler aka DLSS is now main reason people choose NVIDIA. Games now tends to non-removable upscaler. That what AMD need

AMD always "needs" the thing Nvidia has, won't make a difference to sales though.

 

[igor_kavinski]

AMD always "needs" the thing Nvidia has, won't make a difference to sales though.

They suffer because they try to play catch up. Be original. Blow minds. Create real desire.

 

[SolidQ]

AMD always "needs" the thing Nvidia has, won't make a difference to sales though

If AMD will have same quality upscaling and good RT, then choose between AMD and NV gonna be just price to performance

 

[soresu]

They suffer because they try to play catch up. Be original. Blow minds. Create real desire.

Yes and no.

They suffer more because most/all of their effort is open.

nVidia 'originality' exists solely to create their own Apple style walled garden of features and services inside the PC gaming market, to the end point of making it the de facto nVidia gaming market.

ie AMD's efforts are more about mitigating a possible nVidia stranglehold monopoly on the entire PC gaming market, which probably would have happened already if not for nVidia having become so detestable to the 2 hi end game console manufacturers that they stopped using them as a gfx IP supplier in the PS4/XB1 generation.

Until pretty recently it has been entirely on AMD to keep that from happening in the PC market by offering an open source option - with Intel around now it is significantly harder for nVidia to keep trying this as Intel have deeper pockets (for now at least) to keep them from tying it all up for themselves.

 

[inquiss]

They suffer because they try to play catch up. Be original. Blow minds. Create real desire.

Nah by then it will be some other feature that Nvidia is perceived to be marginally better in, even when AMD is better in other features it doesn't matter.

 

[Aapje]

Nah by then it will be some other feature that Nvidia is perceived to be marginally better in, even when AMD is better in other features it doesn't matter.

The point of the comment you replied to is that AMD should be first with a feature that Nvidia doesn't have.

 

[inquiss]

The point of the comment you replied to is that AMD should be first with a feature that Nvidia doesn't have.

They do, you just dont remember them because they don't market them as well and because GPU buyers ignore AMD features and hype up Nvidia's. As soon as AMD reaches feature parity or exceeds Nvidia in something, the hype machine just pivots to a different perceived strength of Nvidia and then only the new thing matters. Rinse repeat.

 

[Mahboi]

Exactly. AFMF/AFMF2? Nvidia has nothing like it.
FSR2 being available as open source and thus completely customisable? Go take in the DLSS binary and be quiet about it.
FSR 3 is objectively better than DLSS 3 pretty much across the board. Early frame pacing problems notwithstanding.

There is a general issue with AMD putting out software that is customisable but imperfect, while Nvidia idiot proofs everything and loves to control everything themselves.
I do agree with the NVers that having an AI corrected upscaler is very necessary, but I do not think that it is a sufficient enough argument to claim that "NV wins everything".

Frankly, it's down to a question of general communication competence. If you look at something like FSR, you'd assume you plug n play the binary, not at all. Everyone who touched it circa FSR 2.1 said that it just doesn't work well and you're expected to tune it. This is why you end up with something like FSR 1 being somehow able to run No Man's Sky on Switch pretty flawlessly at 720p, while FSR 2 implementations that aren't done with an FSR expert's touch just kinda look bad.

I think NV's approach is better than AMD's here, most studios won't have any finetuning past the bare minimum nor will they commit someone to become an upscaling expert to tune it to perfection.