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]

From DF in that 5070 review: "The 9070/XT reviews won't be out until tomorrow, so we can't talk about numbers there, but everyone will be talking about about AMD's answer to Nvidia's 50 series lineup, let's just put it that way. And I'm not sure it's going to be particularly great news for team green."

 

[gaav87]

From DF in that 5070 review: "The 9070/XT reviews won't be out until tomorrow, so we can't talk about numbers there, but everyone will be talking about about AMD's answer to Nvidia's 50 series lineup, let's just put it that way. And I'm not sure it's going to be particularly great news for team green."

Watch GN he litteraly said 9070xt fps numbers during 5070 review kinda crazy... both steves close to nda break

 

[SolidQ]

Watch GN he litteraly said 9070xt fps numbers during 5070 review kinda crazy... both steves close to nda break

GN re upload video? There no time stamp anymore

 

[Ghostsonplanets]

9600XT is N44 or N48 that is question

https://twitter.com/x/status/1896921459660112307

N44. I don't see they doing a further N48 cutdown.
9060 XT 16GB - 32 CUs - $349/$329
9060 16GB - 28 CUs - $299
9050 8GB - 28/24CUs $229

That's what I think

Watch GN he litteraly said 9070xt fps numbers during 5070 review kinda crazy... both steves close to nda break

Which minute?

 

[gaav87]

GN re upload video? There no time stamp anymore

No he starts here and litteraly says +- perf of 9070xt after taht on every chart for rt

 

[gaav87]

Here he dont even care to hide it and just say it xD

 

[Hitman928]

No he starts here and litteraly says +- perf of 9070xt after taht on every chart for rt

Sounds like he's saying the 9070xt numbers in reference to what AMD themselves published, not based on his own numbers, though maybe hinting that he's seeing similar numbers.

Here he dont even care to hide it and just say it xD

Here he just points out that the 5070Ti is the 9070XT's target competition. Not sure he's trying to let on to any performance numbers.

 

[gaav87]

Sounds like he's saying the 9070xt numbers in reference to what AMD themselves published, not based on his own numbers, though maybe hinting that he's seeing similar numbers.



Here he just points out that the 5070Ti is the 9070XT's target competition. Not sure he's trying to let on to any performance numbers.

IF you really listen man and watch the whole review xD explain 14:14 then xD

 

[Hitman928]

IF you really listen man and watch the whole review xD explain 14:14 then xD

Sure, that's different than giving actual performance numbers though.

 

[gaav87]

Sure, that's different than giving actual performance numbers though.

He did not look at the results but 9070=7900xt in 4k starfield cmn man... xD

 

[Hitman928]

He did not look at the results but 9070=7900xt in 4k starfield cmn man... xD

I am not going to watch a 35 min video for clues for something that's going to come out tomorrow. But more specifically, this is what you said:

Watch GN he litteraly said 9070xt fps numbers during 5070 review kinda crazy... both steves close to nda break

No he starts here and litteraly says +- perf of 9070xt after taht on every chart for rt

Then you provided timestamps but he never said this. The only thing I heard from what you timestamped that could qualify was 9070 (non-XT) was roughly equal to the 7900XT in 4K Starfield. So yeah, there and elsewhere, he's not so subtly telling everyone to wait for tomorrow's results, but again, that's far different than giving FPS/performance numbers before release tomorrow. If he gave actual numbers or said performance levels elsewhere, great, but I didn't hear it in anything that was timestamped.

tl;dr: I was told there would be early 9070(XT) numbers, but none were found.

 

[Mopetar]

But some people do.
Produce a "symbolic" quantity of BIG chips for those who will surely pay to get some of the "mindshare" back.

It wouldn’t be worth it for AMD unless they were going to commit to producing them at volume. The cost of the engineering hours to lay out a design and the masks to produce it already mean it can't be a symbolic part.

NVidia can sell their big die for truckloads of money into the datacenter market and the defective parts get handed down to the consumer market. If NVidia didn't have this option, it's unlikely they would be going as big as they are.

If AMD eventually creates a unified architecture that can be used in any market they too can make a chungus die that will also act as a flagship gaming GPU. They're better off focusing time and resources on creating an architecture that goes toe to toe with or even exceeds what Nvidia has on offer.

 

[adroc_thurston]

Unlike the fairly stale and nearly tapped out state of raster gfx, the software side of GPU RT/PT is still in its wild west period of discovery, with a lot of potential low hanging fruit remaining to improve performance with the same hardware without using super resolution haxx.

Point of fact, the software side has improved performance many times more than the hardware side since Turing was announced in 2018.

Uh, no? None of that happened.
It's just vendors shipping incrementally better FF slop.

 

[marees]

Sapphire stocks

https://twitter.com/x/status/1896973507013304450

 

[branch_suggestion]

Uh, no? None of that happened.
It's just vendors shipping incrementally better FF slop.

Yeah, Intel did RTRT demos with Larrabee eons ago, DXR is a sad joke, FF nonsense.
AMD is steadily building a RTRT programmable pipeline it seems, basically ImgTec but it actually happens.

 

[SolidQ]

Sapphire stocks

UK also seems fine

UK has "ample" day 1 RX 9070 series GPU stock - retailers confirm - OC3D

AMD's Radeon RX 9070 series will have "ample" day 1 stock in the UK. AMD delivered more supply than any of Nvidia's RTX 50 series launches.

overclock3d.net

 

[soresu]

It's just vendors shipping incrementally better FF slop.

I was talking about ReSTIR and other related techniques.

 

[soresu]

AMD is steadily building a RTRT programmable pipeline it seems, basically ImgTec but it actually happens.

Oh?

 

[adroc_thurston]

I was talking about ReSTIR and other related techniques.

They just make the IQ worse, but perf better.
DXR itself is still stuck in its DX7 era, 7 years after the fact.

Oh?

yeah, pretty obvious given they're sticking to shaders.

 

[soresu]

yeah, pretty obvious given they're sticking to shaders.

Oh? = Please elaborate. (for me at least)

Not questioning the claim.

They just make the IQ worse, but perf better.

Yeah but no.

While ReSTIR techniques certainly do suffer from temporal artifacting there are methods to filter it out, and variants that minimise, if not reduce the artifacting entirely without filtering while suffering a bit of a perf dip.

There are other possible ways of getting to ReSTIR's performance level, but they are basically newborn and untested by comparison.

For the time being ReSTIR is the only viable way of playing a path traced game of any complexity in real time.

As to IQ of ReSTIR in general - the alternative is basically just denoising pure path traced frames, which I can assure you at playable FPS would look unwatchable from the high sampling variance leaving the frame looking like mush after denoising is applied, so I don't see what you are getting at.

Sure ReSTIR isn't the perfect solution, and certainly not acceptable for offline VFX work where unbiased rendering accuracy is the maxim.

But for games running at real time it is fine.

To be clear, it is denoising that makes IQ look worse.

It smears details in the process of reducing the MC sample variance, because it is basically just averaging out pixel data with a dash of heuristics involved.

Last point - I am really not fond of giving praise to nVidia for anything, but credit goes where it is due, ReSTIR is frickin amazing.

 

[Ghostsonplanets]

yeah, pretty obvious given they're sticking to shaders.

I remember some devs complaining about DXR and Nvidia RT being basically a black box they interact with and lamenting that there should be a programmable RT pipeline.

If AMD can pull this off with RDNA 5, considering that's the IP PS6 and Xbox Next are using, it might be a major disruption to NV uArch.

 

[adroc_thurston]

Oh? = Please elaborate. (for me at least)

Not questioning the claim.

Sooner or later MS will vomit a DXR spec that's a rough equivalent to DX9 SM 2.0 and when that time comes, AMD would have the easiest time adopting it since they're sticking to keeping all the tricky RTRT bits on shaders.

Yeah but no.

While ReSTIR techniques certainly do suffer from temporal artifacting there are methods to filter it out, and variants that minimise, if not reduce the artifacting entirely without filtering while suffering a bit of a perf dip.

There are other possible ways of getting to ReSTIR's performance level, but they are basically newborn and untested by comparison.

For the time being ReSTIR is the only viable way of playing a path traced game of any complexity in real time.

As to IQ of ReSTIR in general - the alternative is basically just denoising pure path traced frames, which I can assure you at playable FPS would look unwatchable from the high sampling variance leaving the frame looking like mush after denoising is applied, so I don't see what you are getting at.

Sure ReSTIR isn't the perfect solution, and certainly not acceptable for offline VFX work where unbiased rendering accuracy is the maxim.

But for games running at real time it is fine.

no this is basically cope.
the #1 issue is people trying to make complex raytraced scenes work with miserable SPP count.
They ought to add micro- prefix to RTRT since that's what it is pretty much.

I remember some devs complaining about DXR and Nvidia RT being basically a black box they interact with and lamenting that there should be a programmable RT pipeline.

not 'some', sebbbi did. right on xitter.

 

[reaperrr3]

N44. I don't see they doing a further N48 cutdown.
9060 XT 16GB - 32 CUs - $349/$329
9060 16GB - 28 CUs - $299
9050 8GB - 28/24CUs $229

That's what I think

My wish:
the above (not for myself, but for the market/consumers).

My guess:
9060 XT 16GB - 32 CUs - $379
9060 16GB - 28 CUs - $349
9060 8GB - 28 CUs - $279
9050 6GB (96bit) - 24 CUs - $199

I blame upselling shenanigans + memory cost + no NV cards with more than 8GB below the upcoming $449 5060 Ti 16GB.

 

[Tuna-Fish]

I would be surprised if there were two 60 SKUs with 16GB.

 

[Saylick]

Sooner or later MS will vomit a DXR spec that's a rough equivalent to DX9 SM 2.0 and when that time comes, AMD would have the easiest time adopting it since they're sticking to keeping all the tricky RTRT bits on shaders.

no this is basically cope.
the #1 issue is people trying to make complex raytraced scenes work with miserable SPP count.
They ought to add micro- prefix to RTRT since that's what it is pretty much.

not 'some', sebbbi did. right on xitter.

This is interesting. Do you got a link to Seb's post on Xitter? I'd like to read about it.