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]

. I'm considering recent (unreliable) rumors

This Source? And both have difference Die area Anyway maybe Kepler_L2 can hint 200+ mm or 300+mm, or both wrong. As i know he's saying there no matrix hardware, so both "Sources" seems miss it

Spoiler

 

[jpiniero]

Do keep in mind that N4P is not much of a shrink compared to N5 and MCD even wouldn't be that much of a shrink either because of the mediocre cache scaling.

A theoretical monolithic N32 on N4P would still be quite big. Depending on how much die space is being taken up by the connection logic.

Course being MLID and RGT, they are completely making stuff up so I wouldn't sweat it too much.

 

[SolidQ]

A theoretical monolithic N32 on N4P would still be quite big.

Big question is hiding here. How much is difference compare to RDNA3, ipc etc

 

[jpiniero]

The legit rumor was that they greatly improved the RT. Figure if that's true, that is going to take up a decent amount of space you would have to think.

 

[SolidQ]

The legit rumor was that they greatly improved the RT.

Maybe kepler can hint us, about RT/mm/Wgp

 

[soresu]

Not sure if these "leaks" are the kind that you like to hear

I'm just sick of the endlessly repetitive content spewing machines that those channels produce based on what is basically the same info just slightly changed, or sometimes just the same but presented differently.

It's basically just a way to game the algorithm when making video content, and they pad the runtime to try and get some of that YT Premium income too.

I've seen numerous channels that do the same thing using different subject matter like 'woke' media in Hollywood, or scitech news - all subjects where they can go over and over the same material with slight changes and pad the runtime to 10-20 minutes a pop so that they can make a living off of it.

 

[Mopetar]

Part of it is YouTube's fault. The algorithm for their recommendation system is less likely to show you videos for channels that don't do those things.

It's also our fault as consumers for clicking on them. When there's a market for this slip, is it all that surprising that people rush in to provide it?

 

[coercitiv]

It's also our fault as consumers for clicking on them.

The platform is built and optimized to minimize choice. "Our fault" is we're still using the product. I've given up on Google Search many years now, so I was kinda' surprised to see folks complain about it's quality spiraling down. (by folks I mean journalists and IT professionals) I still use Google Maps though, and last weekend I got an ad placement instead of a destination I searched for on my phone. I guess it's time to dump Maps as well. The enshittification is real and we're living it.

 

[blckgrffn]

The platform is built and optimized to minimize choice. "Our fault" is we're still using the product. I've given up on Google Search for many years now, so I was kinda' surprised to see folks complain about it's quality spiraling down. (by folks I mean journalists and IT professionals) I still use Google Maps though, and last weekend I got an ad placement instead of a destination I searched for on my phone. I guess it's time to dump Maps as well. The enshittification is real and we're living it.

I "fight" my kids so hard on YT usage. There are some folks who have spent a lot of time creating fun StarCraft videos (the cartooned ones are amazing), replays of Pro matches and some stuff that is really "entertainment". Then there is so much creepy BS "This video will make you fart!" and the algo is so good that just being there will harpoon you into sticking around. My dad sits and scrolls FB videos for like... hours. Ugh. I can't blame them because its a honey pot, but man...

 

[SolidQ]

So RDNA4 RT? 60CU and for now 2,18ghz

 

[GodisanAtheist]

So RDNA4 RT? 60CU and for now 2,18ghz

-That feels like a really big jump in performance for a mid-gen refresh. What good is all that extra headroom when it all has to be crippled for PS5 Basic anyhow?

 

[Aapje]

-That feels like a really big jump in performance for a mid-gen refresh. What good is all that extra headroom when it all has to be crippled for PS5 Basic anyhow?

Apparently this is based on RDNA3 or 4, so has the same dual issue WGPs. So I predict that we get the same huge deflation when RDNA3 launched and all those theoretial FLOPs actually flopped hard.

 

[jpiniero]

MLID tho.

 

[blckgrffn]

-That feels like a really big jump in performance for a mid-gen refresh. What good is all that extra headroom when it all has to be crippled for PS5 Basic anyhow?

To me it does 2 things.

Definitively makes the PS 5 Pro the most powerful console of the generation. Since MS has indicated they aren't making a Series "Even X'ier" version of this console gen from what I have read, this is worth... what it's worth.

Gets early adopters to open their wallets again (I don't get a lot of time to play, but I like it being as Pro as possible, personally )

It's a long term PS6 dev bed, also in my estimation. It's likely that this will allow devs to see if adding RT to their games can add value. That's where the jump is and I guess we'll see how that's used to see how much emphasis PS6 has on it as well. It could be the PS6 differentiator is CPU & other technologies and this GPU could end up being closer to the PS6 GPU than the base PS5. ¯\_(ツ)_/¯ I also could be super wrong.

Apparently this is based on RDNA3 or 4, so has the same dual issue WGPs. So I predict that we get the same huge deflation when RDNA3 launched and all those theoretial FLOPs actually flopped hard.

From what I have read, the hardware RT units are pulled in from RDNA 4 and yeah, the rest is some sort of warmed over/Sony spec RDNA 3. 7800XT with next gen RT units.

 

[Aapje]

From what I have read, the hardware RT units are pulled in from RDNA 4 and yeah, the rest is some sort of warmed over/Sony spec RDNA 3. 7800XT with next gen RT units.

7800 XT level of performance seems extremely optimistic. I'd say around 7700 XT, but then with better RT.

 

[SolidQ]

MLID tho.

it's from Document, even Kepler not decline this

RT units are pulled in from RDNA 4

Seems also WMMA

7800 XT level of performance seems extremely optimistic.

From TPU. More like between 7600 and 6700XT based on DF last video. We have info ~45% UNOPTIMIZED raster perfomance
You can compare Raster and RT

Spoiler

Since MS has indicated they aren't making a Series "Even X'ier

They definitely need Xbox Series E, and that will good fit with 3 letters

 

[soresu]

So RDNA4 RT? 60CU and for now 2,18ghz

Bearing in mind the TFLOPS count is going by the RDNA3 dual issue CU thing.

So divide by 2 as an assumption of real world compute perf baseline, then anything beyond that is a nice surprise.

 

[gdansk]

It's also FP16 and RDNA3 has double rate FP16 so that's about 7700 XT level.

 

[blckgrffn]

It's also FP16 and RDNA3 has double rate FP16 so that's about 7700 XT level.

Keeping in mind that the PS5 is nearly identical spec wise to the 6700 10GB, +45% to the 6700 is likely in the general region of the 7700xt based on my mental benchmark library

I mean the leaks from a while back that were basing it at 7800XT levels from the Pro Dev kits are/were in the ballpark at least. A couple years after release all these differences start to seem vanishingly small.

And at least Sony is usually dedicated to extracting the performance available from the architecture. Whatever is there should get used.

 

[blckgrffn]

it's from Document, even Kepler not decline this


Seems also WMMA

From TPU. More like between 7600 and 6700XT based on DF last video. We have info ~45% UNOPTIMIZED raster perfomance
You can compare Raster and RT

Spoiler

They definitely need Xbox Series E, and that will good fit with 3 letters

They’d probably already have a series 3 but some funny man already played that joke on us

 

[soresu]

And at least Sony is usually dedicated to extracting the performance available from the architecture. Whatever is there should get used.

This has always been the case.

Some of the last generation of games on each successive Playstation look insanely good compared to the first gen.

Just compare The Last Of Us and GTA5 to what the PS3 started with 😅

 

[MrTeal]

Or Tekken Tag Tournament to Tekken 5.

 

[igor_kavinski]

Leak seems to be real: https://www.tomshardware.com/video-...er-rt-performance-and-ai-based-pssr-upscaling

The big news is the up to 4X RT performance increase. I guess that's something really nice and dare I say, LOVELY to expect from RDNA4

 

[TESKATLIPOKA]

The big news is the up to 4X RT performance increase. I guess that's something really nice and dare I say, LOVELY to expect from RDNA4

It's 2-4x as shown in Twitter leak, they have no other info.
And I would take that with a truckload of salt, because we don't know what It exactly means.

PS5: 36CU RDNA2 at 2.23GHz
PS5 Pro: 60CU RDNA3 at 2.18GHz
That's about ~+60% in raster performance, so If I subtract It what will be left is 25-150% higher RT performance.
The difference between RDNA3 vs Ada is only ~40% in Cyberpunk, so we can hardly expect anything above that let alone 150% higher FPS.

 

[igor_kavinski]

The difference between RDNA3 vs Ada is only ~40% in Cyberpunk, so we can hardly expect anything above that let alone 150% higher FPS.

It's not DirectX 12 RT with PS5. They must've implemented something proprietary and optimized to boost RT only for their platform.