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

marees · Aug 31, 2024

soresu said:
Hopefully with N31 being such a conservative CU increase we will see at least 1.33-1.5x more CUs on the RDNA5 hi end, possibly higher still for the enthusiast SKU if chiplets makes a super hi CU count possible without exploding overheads.

I hope there is a monolithic 6700xt/navi 48 style RDNA 5 GPU that could beat 7900xtx. I am scared after AMD nixxed many chiplet designs such as

N36 = 4090 competitor
N43 = original multi chip N48
N42 = N31 replacement
N41 = 4090 killer
N40? = 5090 killer

If the chiplet style RDNA 5 design is a success then we might see numbers such as HD 9990 or even a HD 9999 😉

marees · Aug 31, 2024

marees said:
If AMD launched these 4 SKUs today, then they could price them exactly at

$600 (4070 ti super)
$500
$400
$300 (4060 ti 8gb)

!!!

https://twitter.com/x/status/1828492927523545233

Updated now with infinity cache. (No surprises here. It is coupled to VRAM)

https://twitter.com/x/status/1829610290260164973

marees · Aug 31, 2024

marees said:
So what are known rdna 4 codes?

N48 XTX
N44 XL

What else ?

https://twitter.com/x/status/1829813911115481416

5 RDNA 4 cards, says RGT

TESKATLIPOKA · Aug 31, 2024

soresu said:
Thought RDNA1 was just a single chip?

2

SolidQ · Aug 31, 2024

marees said:
N48 XTX
N44 XL

Seems cutdown N44 is real.
I remember how good was Radeon X800XL price to perfomance

marees · Aug 31, 2024

SolidQ said:
Seems cutdown N44 is real.
I remember how good was Radeon X800XL price to perfomance

RX 8500xt for $250 ??
RX 8600xt for $300 (~ 4060 ti 8gb)

RX 8700xt for $400
RX 8800xt for $500
RX 8800xtx for $600 ? (~ 4070 ti super)

SolidQ · Aug 31, 2024

marees said:
RX 8600xt for $300 (~ 4060 ti 8gb)

Isn't this between 4060ti and 7700?

marees · Aug 31, 2024

SolidQ said:
Isn't this between 4060ti and 7700?

Rt same as 4060 ti is my expectation

Mahboi · Aug 31, 2024

marees said:
So what are known rdna 4 codes?

N48 XTX
N44 XL

What else ?

https://twitter.com/x/status/1829813911115481416

You know it's time to get on with your diet when a 130mm² die is considered "XL".

linkgoron · Aug 31, 2024

marees said:
5 RDNA 4 cards, says RGT

View attachment 106586

I really hope they'll use 8700XT instead of 8800XT. The top Navi will surely be competitive with the 5070 at most. It'll also fit with RDNA1's naming scheme.

maddie · Aug 31, 2024

Mahboi said:
You know it's time to get on with your diet when a 130mm² die is considered "XL".

How is 130mm² relevant? I always thought that XL was a cut die no matter the size.

Timorous · Aug 31, 2024

I would start with 8700XT and move the x700 range to be 16GB parts. I would move x600 to 12GB parts so the 8600XT gets the twice cut 12GB N48 variant.

For N44 I would have 8500XT and 8400XT for the full and cut parts.

While AMD could go with 8800XT for the top part realistically it won't even match the 5070 so I don't think AMD would want their x800 parts priced at the level they would need to price them at to sell as it will be setting up a large price increase for RDNA 5.

marees · Aug 31, 2024

Timorous said:
I would start with 8700XT and move the x700 range to be 16GB parts. I would move x600 to 12GB parts so the 8600XT gets the twice cut 12GB N48 variant.

For N44 I would have 8500XT and 8400XT for the full and cut parts.

Agree. High vram is the only PR win for AMD today

For ex: in BMW once you enable cinematic mode at 1080p 75% upscaled (with FG) it cripples the 1% lows of the 8gb 4060 ti card. While the 16gb 4060 ti card provides a steady 'cinematic fps' (24 fps 1% low & 30fps avg before FG)

So even nv fans acknowledge that 8gb cards cannot play 1080p max

It would be a bad PR move to have 12gb on 8700 class card
Correspondingly if the 8600xt is 12gb, it contrasts well vs the expected 8gb 5060 (5060 ti could be 12gb)

jpiniero · Aug 31, 2024

Timorous said:
While AMD could go with 8800XT for the top part realistically it won't even match the 5070

I am not so sure in raster. Granted the 5070 will definitely cost more.

marees said:
It would be a bad PR move to have 12gb on 8700 class card

Given how things have gone with AMD, it'd be a bigger bad PR move for the 8700 XT to be substantially more expensive than the current 7700 XT price ($389). Which it would be if it was a 16 GB model.

marees · Aug 31, 2024

What do you folks think will be RT performance of navi 48 xtx ?

I am hoping 4070 ti super
others are saying it will definitely be less than that

https://twitter.com/x/status/1829996660988264875

reaperrr3 · Aug 31, 2024

Timorous said:
While AMD could go with 8800XT for the top part realistically it won't even match the 5070

GB205 allegedly has only 50 SM, only 4 more than are active on a 4070 and less than a Super.

Considering how NV usually operates, I wouldn't expect more than 48 enabled SM on a mere 5070.

Meanwhile, N48 not only has 2 WGP/4 CU more than N32, but - perhaps more importantly - 4 SE instead of 3, presumably 128 instead of 96 ROPs, in addition to RDNA4's general architectural improvements.
It'll also clock a good 20% higher than the 7800XT, if latest rumors are true.

In other words, Blackwell would have to beat RDNA4 by wider margin in terms of IPC per SM and/or clocks than Ada vs. RDNA3. Not impossible, but I wouldn't expect GB205 to beat N48 in Raster by a wide margin, if at all.

marees said:
What do you folks think will be RT performance of navi 48 xtx ?

I am hoping 4070 ti super
others are saying it will definitely be less than that

https://twitter.com/x/status/1829996660988264875

50% above a 7800XT would be roughly 79XTX/4070Ti territory.

But the PS5Pro is more like RDNA3.7, has only 2 SE and either less or even no Infinity Cache, so older Raster-uArch and other weaker parts than N48.
If the PS5Pro is 50% above a 7800XT in RT, I'd expect N48XT(X) to be more like 70% above it, which would be pretty much exactly 4070Ti Super territory.

marees · Sep 1, 2024

reaperrr3 said:
50% above a 7800XT would be roughly 79XTX/4070Ti territory.

But the PS5Pro is more like RDNA3.7, has only 2 SE and either less or even no Infinity Cache, so older Raster-uArch and other weaker parts than N48.
If the PS5Pro is 50% above a 7800XT in RT, I'd expect N48XT(X) to be more like 70% above it, which would be pretty much exactly 4070Ti Super territory.

I sincerely hope n48 xtx trades blows with 4070 ti super on RT

That would make an immediate launch for $600 feasible

Otoh, 4070 super territory means they need to price it closer to $500 where it starts cannibalizing the 7900 GRE stocks

In other words if RT is not high then RDNA 4 cannot launch until & unless 7900 GRE stocks are sold out

ToTTenTranz · Sep 1, 2024

marees said:
What do you folks think will be RT performance of navi 48 xtx ?

I am hoping 4070 ti super
others are saying it will definitely be less than that

https://twitter.com/x/status/1829996660988264875

I think the poll isn't very useful because the base premise is wrong.
The PS5 shouldn't be 1.5x 7800XT in RT. In scenarios where the RT performance is a big bottleneck, the boost should be a lot more than +50%.

Sony's own docs claim something between 2x and 4x higher RT performance than the PS5.

TESKATLIPOKA · Sep 1, 2024

reaperrr3 said:
50% above a 7800XT would be roughly 79XTX/4070Ti territory.

But the PS5Pro is more like RDNA3.7, has only 2 SE and either less or even no Infinity Cache, so older Raster-uArch and other weaker parts than N48.
If the PS5Pro is 50% above a 7800XT in RT, I'd expect N48XT(X) to be more like 70% above it, which would be pretty much exactly 4070Ti Super territory.

PS5Pro 50% above 7800XT in RT with similar specs?

That would mean in a real game like A Plague Tale Requiem this:
RTX 4090: 98.8FPS at 4K, with RT enabled only 79FPS (-20% compared to raster).
RTX 4070 Ti super: 62.6FPS at 4K, with RT enabled only 50.1FPS (-20% compared to raster).
7900XTX: 78.6FPS at 4K, with RT enabled only 45.6FPS (-42% compared to raster).
7800XT: 48.8 FPS at 4K, with RT enabled only 28.8FPS (-41% compared to raster).
PS5 Pro or comparable desktop equivalent: 48.8 FPS at 4K in A Plague Tale Requiem, with RT enabled 45.6FPS (-7% compared to raster).

This is simply impossible, even RTX 4090 loses 20% of It's performance after enabling RT while this PS5 Pro would loose only 7%.
Even If AMD managed only 20% reduction in perf, this PS5 Pro would still need to achieve 57FPS without RT, in other words 17% higher FPS to be on par with RX 7900XTX in RT.

gaav87 · Sep 1, 2024

SKU1 (8800xtx) 64CU 256bit 64MB Infinity Cache 16GB ddr6 20Gbps
SKU2 (8800xt) 56CU 256bit 64MB IC 16GB ddr6 18Gbps
SKU3 (8700xt) 48CU? 192bit 48MB IC 12GB ddr6 19Gbps
SKU4 (8600xt) 32CU 128bit 32MB IC 8GB ddr6 18Gbps

SKU1 (8800xtx) 50+ TFLOPs of FP32 (for 50+ must be 3.05GHz or above) [7900xt ~52.5TFLOPs of FP32]
I estimate the compute perf +/- 5% ~7900xt Gaming performance between 7900xt and 7900xtx (closer to 7900xt)
SKU2 (8800xt) Gaming perf 7900 gre
SKU3 (8700xt) Gaming perf 7800 xt
SKU4 (8600xt) performance ~7700 xt

Mahboi · Sep 1, 2024

In hindsight, since there is nothing new to expect in performance, there is ONE thing I really want from RDNA 4. It's encoding quality.
I can do with the rest just fine. Upscale for 4K, accept that it's more or less aiming for good 1440p. But I still watch a lot of streams, and every time, I keep thinking that if you genuinely asked me to market AMD cards to streamers, I couldn't. The encoding quality for h264 is just not worth it. Ofc one can, and should, and I sure do, blame Twitch, since YT has managed AV1 streaming for months/over a year now.
But even for AV1, it's close, but below in a significant enough fashion. It typically doesn't show too badly on the video quality itself, which if you pump enough bitrate can just do it with brute force, but it's visible when almost every recording is somewhere around 10% bigger on AMD than Nvidia IIRC.

I don't know just how impactful a bunch of streamers are, but they unsurprisingly get a lot of visibility. It's kind of an abysmal market image when literally 99% of them have Nvidia.
I'm not holding any hopes for AMD h264 to get good now. Unless somehow they integrate Xilinx IP and the Xilinx IP turns out to be monstrously better than the historical AMD one.
But for AV1, it wasn't THAT far. If someone could kindly present their foot's tip at the Twitch CEO's anus and push as hard as he can inside, we might actually get AV1 before RDNA 5 comes out.

I'm just really hoping for that and a new gen of FSR that has AI image correction like NV. If we can get that, we really are only behind in driver velocity (look at how long it took for RDNA 3 to more or less fully fix VR? 7 months, and then people had to wait for Meta to get things going on their SW too?) and in RT. I still think most people don't care that much for RT. But FSR and better encoding, that's my real hopium.

Kepler_L2 · Sep 1, 2024

gaav87 said:
SKU4 (8600xt) performance ~7700 xt

No way N44 can reach that level of performance.

marees · Sep 2, 2024

Kepler_L2 said:
No way N44 can reach that level of performance.

I am hoping it can hit 4060 ti 8gb level of RT performance. MLID has supplied lots of Hopium

soresu · Sep 2, 2024

tbh I'm all but disinterested in the gaming side of things and just want to see how much Blender Cycles perf increases.

Even on the Eevee side of things it's great because the big overhaul seems to have implemented ReSTIR like ray tracing, bringing the potential quality of Eevee much closer to that of Cycles.

soresu · Sep 2, 2024

Mahboi said:
If someone could kindly present their foot's tip at the Twitch CEO's anus and push as hard as he can inside, we might actually get AV1 before RDNA 5 comes out.

Ironically Twitch actually made a major contribution to AV1 development in the form of S-Frames, which makes their slow transition all the stranger.

Certainly at this point AV1 hw decoding access is fairly widespread, and becoming ever more so since Qualcomm and Apple finally took the stick outta their posteriors and implemented it.

Even where it isn't there on older generations there is plenty of CPU power to decode it efficiently using dav1d, which at this point is quite possibly the most well optimised video decoder library in history.

I don't think that a Vulkan GPU implementation would be a significant improvement for perf/watt over CPU given how complex AV1 is as a codec.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Senior member

Senior member

Platinum Member

Senior member

Senior member

Senior member

Senior member

Golden Member

Platinum Member

Diamond Member

Golden Member

Senior member

Lifer

Senior member

Junior Member

Senior member

Member

Platinum Member

Member

Golden Member

Senior member

Senior member

Diamond Member

Diamond Member