Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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 · Sep 16, 2024

ToTTenTranz said:
is there a chance we'll see FSR4 on RDNA4 dGPUs?

that gonna big fail, if only for APU's, but anyway modders will fix fast it, and gonna work with any dlss game

ToTTenTranz · Sep 16, 2024

Kepler_L2 said:
Probably for anything that supports WMMA.

This is indeed great if it even supports Phoenix SoCs.

marees · Sep 18, 2024

As per early previews PSSR is (almost) as good as DLSS/XeSS (Except for a few temporal instability)

Will FSR 4 be as good as PSSR
Will hardware of RDNA 4 support FSR 4 or would it need XDNA from strix / krackan
Will AMD delay RDNA 4 so that it doesn't steal Sony's thunder (PSSR & PS5 pro)

marees said:
The early verdict is in:

PS5 pro is good & Sony delivers on its promise (almost)

PSSR delivers quality mode visuals (almost) with performance mode fps (slightly better. The pro locks better to 60 fps)

PSSR works its magic. But very slightly behind DLSS/XeSS ( in temporal stability)

if you've been waiting to play Final Fantasy 7 Rebirth, it may be worth waiting a couple months longer, because PS5 Pro seems like the best place to play the Square-Enix epic by a long shot.

https://www.eurogamer.net/digitalfo...vers-huge-image-quality-improvements-at-60fps

Pro vs Quality mode:
the biggest issue here is that the PS5 Pro just isn't that temporally stable relative to the base machine in its graphics mode. In shots without a lot of camera movement, it's fairly easy to find quite a bit of breakup in the image, at least relative to the graphics mode. The image stability overall is more similar to the old performance mode. I don't think this is anything like a dealbreaker here, but the general instability of the image is the biggest sticking point if you are measuring it up against its full-res PS5 counterpart. Here, solutions like DLSS and XeSS typically deliver more temporally stable results. The brief snippets of Ratchet and Clank that we have appear to exhibit a similar problem, suggesting this may be a characteristic weakness of the current version of PSSR.

Overall, I think image quality is similarly good between the two consoles when comparing base PS5 quality mode to the 60fps Pro footage we have, although they can also look fairly different in some areas. The Pro definitely gives you greater image detail, but it comes at the cost of more temporal instability

Pro vs Performance mode:
The intro cutscene shows off the impressive fidelity the PS5 Pro is capable of, delivering a crisp, sharp 4K image - a far cry from the blurry and imprecise 1080p-like rendering featured on the base console at the same 60fps update.

FF7 Rebirth packs a lot of geometry into every frame, boasting primitive shader-powered polygonal detail, and the PS5 Pro simply handles this much better than the base console. Ghosting on camera movement is much harder to spot on the Pro as well. Combat frustrates the base machine a little, with blurry particle effects and poor image clarity producing a muddy look. PS5 Pro handles this scene impressively - much better than I was expecting - resolving crisp 4K-like detail even in relatively fast combat scenes. The motion blur looks quite pristine and is relatively artifact free. Particle effects look generally good in motion, though some freeze-frames reveal a slight chunkiness relative to the rest of the presentation.

The base resolution is very similar to the resolution range of the base game, which hit comparable resolution figures in its performance mode. However, the base console didn't use any form of temporal upsampling, instead taking that final rendered image and simply upscaling it, perhaps using a bilinear or bicubic upscaling solution. The game's final resolve also appeared oddly soft, lacking detail relative to some other games with similar resolutions. Square-Enix also offered a 'performance-sharp' mode, which used a different upscaling method that produced sharper results, albeit with additional blockiness. Preferences here are a matter of personal taste, but neither option offered the kind of image clarity that most other 60fps games on consoles are capable of delivering.

soresu · Sep 18, 2024

marees said:
As per early previews PSSR is (almost) as good as DLSS/XeSS (Except for a few temporal instability)

Will FSR 4 be as good as PSSR

Will hardware of RDNA 4 support FSR 4 or would it need XDNA from strix / krackan

Will AMD delay RDNA 4 so that it doesn't steal Sony's thunder (PSSR & PS5 pro)

The likelihood is that PSSR like FSR and DLSS will only improve over time as Sony do quality tuning.

I highly doubt that AMD would delay RDNA4 over Sony's plans.

Any delay at this point is about volume of older product in the market.

Delaying Strix Halo until PS5 Pro has had some time to sell it is another matter, as SFF/UCFF PCs built with the new APU SKU have the potential to exceed base PS5 specs and then some, making it a less viable choice without the mid cycle console upgrade.

Using XDNA for FSR4 seems unlikely unless integration is extremely tight with latency between NPU and GPU so low that it doesn't bite into the upscalers perf advantage.

I just don't see that being viable as it would have to compute the lower res frame, copy the data to the NPU, upscale and then copy the data back to the GPU to display it.

That's not even getting into the power wasted to transfer frames to and from the NPU vs keeping them in the GPU WGPs, which might mitigate any power efficiency advantage from using XDNA's inference engine.

SolidQ · Sep 18, 2024

marees said:
As per early previews PSSR is (almost) as good as DLSS

PSSR gonna be better, at least it's massive better compared to DLSS1

Will hardware of RDNA 4 support FSR 4

Even if it not goind to RDNA4, modders will fix it

GodisanAtheist · Sep 18, 2024

Mahboi · Sep 18, 2024

And to be fair, I don't mind at all. The higher the base res, the more functional software res is. I don't think 8K was ever going to be worth amount of compute units. Upscale anything above 4K from 4K and that'll do nicely I think.

GodisanAtheist · Sep 18, 2024

Mahboi said:
And to be fair, I don't mind at all. The higher the base res, the more functional software res is. I don't think 8K was ever going to be worth amount of compute units. Upscale anything above 4K from 4K and that'll do nicely I think.

- Yeah, my position on upscalers has softened over time, I do think they have a solid use case in ultra-high res scenarios and in keeping old hardware relevant well passed its typical expiry date (my use case with FSR and my now ancient 980Ti, still playing Hogwarts Legacy and other "new" UE4 games at 30-60FPS with FSR).

If you need upscaling on brand new hardware for common resolutions like 1080P and 1440P then something has gone horribly wrong with your performance targets.

Thunder 57 · Sep 18, 2024

GodisanAtheist said:
If you need upscaling on brand new hardware for common resolutions like 1080P and 1440P then something has gone horribly wrong with your performance targets.

Hasn't that already begun happening though?

GodisanAtheist · Sep 18, 2024

Thunder 57 said:
Hasn't that already begun happening though?

- Yep, and that just means I wait even longer to play a game until I have the hardware that can run it at native res.

static shock · Sep 20, 2024

I'm just thinking that it will be at N3E.

jpiniero · Sep 20, 2024

static shock said:
I'm just thinking that it will be at N3E.

Not sure what product you are talking about but it'll be N4.

static shock · Sep 20, 2024

jpiniero said:
Not sure what product you are talking about but it'll be N4.

Nevermind.

static shock · Sep 20, 2024

jpiniero said:
Not sure what product you are talking about but it'll be N4.

N4 for gaming? TSMC does have good yields now, and AMD will IMO not fab more than 300mm2 of a piece of silicon for the fastest and biggest GPU. This remember me of HD4850 strategy. Well i'm still think that i'm talking something that everybody knows.

static shock · Sep 20, 2024

I will shot one: my bet is a 2.5 performance per watt vs Rdna3, and N3E. Performance around RX7900xt for the fastest Videocard.

gdansk · Sep 20, 2024

static shock said:
I will shot one: my bet is a 2.5 performance per watt vs Rdna3, and N3E. Performance around RX7900xt for the fastest Videocard.

2.5x performance per watt and around 7900XT performance would mean it's about 120W TGP.
May want to reconsider your spitballing.

static shock · Sep 20, 2024

gdansk said:
2.5x performance per watt and around 7900XT performance would mean it's about 120W TGP.
May want to reconsider your spitballing.

Why reconsider? You never go to YouTube and check the gpupwr for an NVada card? Amd haves a chance to make something that is top-notch and/or perfectly crafted for GPU too. Remember 8800 gtx? Then remember Fermi v1 and v2(500 series)? And hd 4850?

darkswordsman17 · Sep 20, 2024

static shock said:
Why reconsider? You never go to YouTube and check the gpupwr for an NVada card? Amd haves a chance to make something that is top-notch and/or perfectly crafted for GPU too. Remember 8800 gtx? Then remember Fermi v1 and v2(500 series)? And hd 4850?

Sorry but you're delusional and ignoring that sure, such gains used to be feasible just simply from process (which you seem to think is gonna be N3 but that is almost certainly not going to be the case). And even if there were ability to make such leaps via architecture, an established company isn't going to do that because it will bring software compatibility issues as it would require substantial reworking of their architecture. This is where I'm assuming you'd point to RDNA 1 and 2, but also would ignore that is also what I'm talking about as it was AMD reworking their GPU substantially, removing a lot of compute focused aspects to tailor it more to gaming (there's still compute, its just not outsized aspect as it was in GCN era). Plus it also brought the software issues I was talking about. It ultimately settled down, and it looked like "AMD fine wine" as things improved (especially RDNA2 which benefitted from the work that had to be done to improve things after RDNA1, but I believe it also had a pretty long period where they reworked the driver and it caused a lot of issues - think this happened after launch but before RDNA3? Or maybe that was between RDNA 1 and 2?)

RDNA4 is simply AMD pairing things down to make a modest and efficient GPU. Ditching chiplets which hurt efficiency (especially idle and low power) with not that much benefit to die size or other like heat distribution. Which sounds like we could see that level of gains in ray tracing performance, just because of how lacking AMD's current and previous gens have been in that regard.

Maybe I'm wrong, but I think AMD has already said RDNA4 will be N4. I don't think any rumors have pointed to them delaying it to put it on N3. And I think for cost reasons alone it likely is not gonna be N3.

static shock · Sep 20, 2024

darkswordsman17 said:
Sorry but you're delusional and ignoring that sure, such gains used to be feasible just simply from process (which you seem to think is gonna be N3 but that is almost certainly not going to be the case). And even if there were ability to make such leaps via architecture, an established company isn't going to do that because it will bring software compatibility issues as it would require substantial reworking of their architecture. This is where I'm assuming you'd point to RDNA 1 and 2, but also would ignore that is also what I'm talking about as it was AMD reworking their GPU substantially, removing a lot of compute focused aspects to tailor it more to gaming (there's still compute, its just not outsized aspect as it was in GCN era). Plus it also brought the software issues I was talking about. It ultimately settled down, and it looked like "AMD fine wine" as things improved (especially RDNA2 which benefitted from the work that had to be done to improve things after RDNA1, but I believe it also had a pretty long period where they reworked the driver and it caused a lot of issues - think this happened after launch but before RDNA3? Or maybe that was between RDNA 1 and 2?)

RDNA4 is simply AMD pairing things down to make a modest and efficient GPU. Ditching chiplets which hurt efficiency (especially idle and low power) with not that much benefit to die size or other like heat distribution. Which sounds like we could see that level of gains in ray tracing performance, just because of how lacking AMD's current and previous gens have been in that regard.

Maybe I'm wrong, but I think AMD has already said RDNA4 will be N4. I don't think any rumors have pointed to them delaying it to put it on N3. And I think for cost reasons alone it likely is not gonna be N3.

And your guess about power efficiency gains? At N4? Amd is doing this to achieve a landmark performance per watt uplift. I viewed on a forum that the RT engine is at Ada Lovelace levels. Chiplets centainly rock, but the video card business is, in the end, about how much they can build a monster efficient GPU, and chiplets are not there in consoles, handhelds, AI acellerators, workstations and datacenters and notebooks.

Anyway a n3 GCD plus n4 MCDs willbe not a bad idea. Chiplets do cards with more affordable prices. Remember N3 yields are already very good for a 200mm2 GCD.

adroc_thurston · Sep 21, 2024

darkswordsman17 said:
Ditching chiplets which hurt efficiency (especially idle and low power) with not that much benefit to die size or other like heat distribution.

They're not really ditching them, just that mainstream parts won't be tiled for some while.

MarkPost · Sep 23, 2024

So is there any consensus about a release timeframe?

gdansk · Sep 23, 2024

MarkPost said:
So is there any consensus about a release timeframe?

No. All the software support and marketing talk from AMD suggests sooner than later. But partners and inventory suggest later.

poke01 · Sep 23, 2024

All want is better RT from RDNA4. I could then use that card in Linux with no issues and also with Blender

MarkPost · Sep 23, 2024

gdansk said:
No. All the software support and marketing talk from AMD suggests sooner than later. But partners and inventory suggest later.

thanks

soresu · Sep 23, 2024

poke01 said:
All want is better RT from RDNA4. I could then use that card in Linux with no issues and also with Blender

Same - ironically the news just dropped about HIP RT v2.4 so it's good to know that AMD is still throttle forward on that.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Member

Senior member

Diamond Member

Senior member

Diamond Member

Golden Member

Diamond Member

Diamond Member

Diamond Member

Member

Lifer

Member

Member

Member

Diamond Member

Member

Lifer

Member

Diamond Member

Senior member

Diamond Member

Platinum Member

Senior member

Diamond Member