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

 

[soresu]

a) imgtec had RTRT h/w way before NV
b) it's a sorta-natural evolution of gfx.

#1. IMGTEC bought it from Caustic and took forever to capitalise on it - plus Intel were playing around with it even earlier still (I wonder how many people remember that ancient Quake 4 RT demo).

#2. Yes it is a sorta natural evolution.

But not a necessary one for real time gaming where a hacky, overly complex raster gfx game engine that performs well on less powerful hardware is to be preferred over the alternative necessities for real time RT/PT gaming.

It makes far more sense for VFX production than it ever will for games where it benefits the developers far more than the actual users.

It would have been a much more sensible strategy to slowly introduce it to games over time with progressively more advanced features, rather than trying to ram it down peoples throats from day 1 with what was arguably the most idiotic campaign ever that tried to glorify perfectly mirrored surfaces that don't exist in real life simply because even a slight amount of roughness makes reflected ray calculations a whole lot more complicated.

 

[soresu]

Really curious to see whether that Bolt/Zeus GPU thing is just smoke or there is real fire behind it.

 

[inquiss]

Really curious to see whether that Bolt/Zeus GPU thing is just smoke or there is real fire behind it.

It's not going to have any effect on gaming. Maybe it will find it's niche in render farms for film.

 

[soresu]

It's not going to have any effect on gaming. Maybe it will find it's niche in render farms for film.

More in my wheel house anyway, these days my eyesight simply isn't amenable to gaming.

 

[soresu]

Interesting video. I've noticed in different videos, 1% on NV is worse

Lol, modern AAA game Q/A at release, innit swell?



Does anyone notice anything wrong with this image?

Or did this Spidey gain a new powah I don't know about? 😂

 

[Gideon]

Very Interesting interview with Lisa Su

https://twitter.com/x/status/1904418055285002692

- AMD Lisa Su
What we really wanted to do is bring the best gaming capability to a very, let's call it good price points. So that we could get more gamers have access to this technology. Of course, everyone likes a very high end GPU. But it's not so many people can access it.

9070 XT has been a fantastic success. It's the No.1 selling for all of the AMD Radeon generations. For first week sales is the by far 10x higher than previous generations. And we're increasing the manufacturing so that we can have more gamers to have access.

and also

when tony said i really hope you can bring it (strix halo) into deskop.
lisa said yes yes yes of course.

Though probably here she means something akin to Framework Desktop

 

[tsamolotoff]

But not a necessary one for real time gaming

Actually, if you remember the initial RTX presentation back in the days, most of it was devoted to professional applications of RTRT while gaming was on the backseat. I guess nVidia didn't initially plan for using it as the next iteration of gimpworks, but the opportunity naturally presented itself (and now thanks to AI craze they can probably dictate anything to MS in exchange for priority shipments of chips or whatever)

 

[soresu]

most of it was devoted to professional applications of RTRT while gaming was on the backseat

Even real time isn't necessary IMHO for RT in professional use cases.

A nice extra for viewport preview visual accuracy in DCC, but not a necessity.

Accelerating offline rendering on the other hand, that's a killer app to be sure - in this arena many samples per pixel and accuracy (minimal denoising) are king, making real time non viable.

Likewise other applications that simulate rays in professional use cases unrelated to graphics where accuracy is prefered to timely, good enough data.

If AMD finally get their act together and bring HIP RT to something closer to OptiX level maturity + do the coding support to kickstart major commercial renderers (Renderman, Arnold, Karma etc) using it, then I could definitely see myself buying a multi x16 slot mobo to gradually put more GPUs in as I get the funds to do so.

This is IMHO the greatest benefit of GPU offline rendering relative to CPU - you can't just put a 2nd high end server CPU into a single socket consumer mobo, but there are still consumer mobos with up to three x16 PCIe slots, and pro (Threadripper) mobos that go as high as 7.

Very Interesting interview with Lisa Su

https://twitter.com/x/status/1904418055285002692

and also


Though probably here she means something akin to Framework Desktop

I think that this is that interview, it's with ASUS China's manager Tony Yu...

 

[tsamolotoff]

Accelerating offline rendering on the other hand, that's a killer app to be sure

Current RTX 96 gb monsters for sure, but was it enough back in the days for serious professional stuff ? (I'm not familiar with this side of 3d graphix, so no idea how much VRAM is needed for proper offline rendering of ILM grade stuff)

 

[poke01]

If AMD finally get their act together and bring HIP RT to something closer to OptiX level maturity + do the coding support to kickstart major commercial renderers (Renderman, Arnold, Karma etc) using it, then I could definitely see myself buying a multi x16 slot mobo to gradually put more GPUs in as I get the funds to do so.

HIP RT has a long way to go

 

[soresu]

Current RTX 96 gb monsters for sure, but was it enough back in the days for serious professional stuff ?

Most serious renderers have a feature that allows them to stream data in and out of memory during the render to alleviate this problem.

It's not ideal, but the best solution until something like Bolt/Zeus comes along with upgradable memory capacity.

 

[Gideon]

Most serious renderers have a feature that allows them to stream data in and out of memory during the render to alleviate this problem.

It's not ideal, but the best solution until something like Bolt/Zeus comes along with upgradable memory capacity.

I wonder if low end GPUs with LPDDR CUDIMMs wont start to make sense at some point.

 

[adroc_thurston]

with LPDDR CUDIMMs wont start to make sense at some point.

a) no
b) any socketed LPDDR will be very slow

 

[poke01]

a) no
b) any socketed LPDDR will be very slow

i wanna know how much of a scam is Bolt/Zeus then

 

[adroc_thurston]

i wanna know how much of a scam is Bolt/Zeus then

Wow a manycore R-V slop "startup".
Never happened before!

 

[Gideon]

a) no
b) any socketed LPDDR will be very slow

That in itself shouldn't be a blocker for GPUs as they hide latenc very well. But I agree it makes little sense. Particularily to the AIBs who would have one less thing to "optimize" to make money (the memory)

 

[Win2012R2]

If it's not RTRT then what tech will allow us to stop doing Quake-era baked lighting?

Without fully dynamic lighting complex worlds have to be largely static with pre-baked stuff, so a lot of cool destructions are off, what other technology promises to achieve that?

Maybe the way Nvidia did it was not meant to be the best and DXRT is crap implementation that only suits Nvidia's approach, perhaps AMD got better approach, but in any case fundamentally it's the right and probably only direction to get to the next level, even if takes next 10 years to have even peasant consoles supporting it at decent enough level for developers to risk it.

 

[Jan Olšan]

So people claiming to me. only NVIDIA drives the industry forward, the rest companies just copy!

AMD is stalling gaming progress! AMD GPUs slowing down innovation! Gotta get rid of them!

Absolutely nothing to do with Nintendo consoles based on underpowered Nvidia chips forming the turd baseline for game devs.

 

[Vikv1918]

If it's not RTRT then what tech will allow us to stop doing Quake-era baked lighting?

Without fully dynamic lighting complex worlds have to be largely static with pre-baked stuff, so a lot of cool destructions are off, what other technology promises to achieve that?

Maybe the way Nvidia did it was not meant to be the best and DXRT is crap implementation that only suits Nvidia's approach, perhaps AMD got better approach, but in any case fundamentally it's the right and probably only direction to get to the next level, even if takes next 10 years to have even peasant consoles supporting it at decent enough level for developers to risk it.

There are only 2 possible lighting technologies - baked lighting and RTRT

 

[blckgrffn]

AMD is stalling gaming progress! AMD GPUs slowing down innovation! Gotta get rid of them!

Absolutely nothing to do with Nintendo consoles based on underpowered Nvidia chips forming the turd baseline for game devs.

I think the next Switch getting a slow 3050 is a huge step up for them and likely a shackle for everyone else. DLSS/FSR will be the norm for every title for docked play, imho.

It’s hard to find it exciting that this will still be the baseline hardware in 2032 or so.

Anyway. I have to get my son to forward me his screen shot results from bench marks but the 9070 is a cool 80% to 150% faster in roughly the same power envelope as his 6700XT. Setting aside the price considerations (seriously this was a $900 Covid GPU so that’s not going to hurt the 9070) it is a big enough increase with the classic generational leapfrog to be worth while, IMO.

 

[Vikv1918]

I think the next Switch getting a slow 3050 is a huge step up for them and likely a shackle for everyone else. DLSS/FSR will be the norm for every title for docked play, imho.

It’s hard to find it exciting that this will still be the baseline hardware in 2032 or so.

Anyway. I have to get my son to forward me his screen shot results from bench marks but the 9070 is a cool 80% to 150% faster in roughly the same power envelope as his 6700XT. Setting aside the price considerations (seriously this was a $900 Covid GPU so that’s not going to hurt the 9070) it is a big enough increase with the classic generational leapfrog to be worth while, IMO.

I would suggest that Switch, mobile phones, handhelds like Steam Deck and other devices with low end GPUs should probably use something like GI 1.0, a very fast approximation of path tracing. And its perfectly fine for them to do so. The point of video games is to play them, and ultra high end lighting solutions only marginally improve their enjoyment (or even detract from it if the ultra high end solution is just tacked on instead of artistically integrated).

The idea that low settings or medium settings hugely detracts from enjoyment of videogames is just hardware vendor propaganda. And the idea that better visual fidelity is synonymous with "pushing gaming forward" is a complete lie if you think about it for more than a few seconds.

 

[MS_AT]

That in itself shouldn't be a blocker for GPUs as they hide latenc very well. But I agree it makes little sense. Particularily to the AIBs who would have one less thing to "optimize" to make money (the memory)

I think by slow Adroc meant bandwidth will be low. And GPUs need lots of BW.

 

[blckgrffn]

The idea that low settings or medium settings hugely detracts from enjoyment of videogames is just hardware vendor propaganda. And the idea that better visual fidelity is synonymous with "pushing gaming forward" is a complete lie if you think about it for more than a few seconds.

Some games, like Astral Chain which I really enjoyed - to a point - but never finished because its just too jank for me to play at Switch levels of performance, just suffer because there isn't enough juice to fulfill their artistic visions and gameplay dynamics. These thresholds are personal. In 2006 or something maybe my expectations would have been so different it wouldn't have mattered.

More capable hardware can give us a better and more immersive - or even more fully realized - experience.

I don't give a crap about what type of illumination is used, the lighting is a part of the artwork vision the game is creating for me. Honestly, I think we are agreeing more than disagreeing here.

 

[Gideon]

I think by slow Adroc meant bandwidth will be low. And GPUs need lots of BW.

Speaking of the devil:

Micron SOCAMM Memory Powers Next-Gen NVIDIA Grace GB300 Servers

This is the new Micron SOCAMM with LPDDR5X memory that will power next-generation NVIDIA Grace GB300 servers

www.servethehome.com

So it seems SOCAMMs work for next-gen Nvidia Grace GB300 servers:





While it's not classical sockets ( no CAMM modules are) they surely seem replacable. Yeah it won't give GDDR7 level bandwidth, but judging by Strix Halo, M4 Max, etc .... it should be enough for even some mid-range GPUs.

 

[SolidQ]

I think the next Switch getting a slow 3050

DF already tested it, at best RTX 2050, and take two guy hinted, a bit slower than base ps4