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]

Azor yesterday mention a little bump in raster. I'd assuming he mean XTX+5% max

 

[soresu]

Nanite is virtualized geometry, you can't 'accelerate' it.

Eh?

Virtualised geometry is part of it I'm sure.

But I was under the impression that it was also an advanced dynamic LOD compute engine that essentially takes raw geometry (and more recently surface displacement) and creates a new LOD based on view distance and pixel resolution so that for sufficiently high poly geometry you would have just one poly (or tri?) per pixel, thus minimalising the amount of geometry to be ray cast against or shaded.

That much at the very least should be possible to accelerate with mesh shaders, primitive shaders etc.

I'm fairly sure that Epic specifically mentioned AMD's primitive shaders on RDNA2 shortly after the initial Lumen/Nanite demo back in 2020, in reference to it being rendered using the PS5.

 

[gaav87]

Azor yesterday mention a little bump in raster. I'd assuming he mean XTX+5% max

I tamed my train. Even if +-5% for both at least it gives normies estimates and they can go to TPU click 6700xt and see +- where they land. I know old TPU relative chart is not accurate.

 

[reaperrr3]

I tamed my train. Even if +-5% for both at least it gives normies estimates and they can go to TPU click 6700xt and see +- where they land. I know old TPU relative chart is not accurate.

If the XT manages that level of performance at stock (~265W), I don't see AMD making the 9070 that slow.

9070 will likely have 8 CUs less, but Frontend (SEs/Primitive Pipes), Backend (ROPs) and Caches should all stay the same, so perf/clock will definitely NOT go down by 12.5% compared to the XT.
And it likely has only 10% less mem bandwidth, too.

Plus, I just don't see AMD making the 9070 only match the GRE and lose to the 4070Ti and 5070 by a hair if AMD could make it roughly tie the 2 NV cards with just a little more TDP/clockspeed.

In short, if the 9070XT placement is accurate, I don't see the 9070 land below 65% on that chart, I'd expect it to land around 67 +/-2% points. AMD will probably not want to price it much more than 10% below the XT anyway, so they can't make it much more than 10% slower, regardless of upselling shenanigans.

 

[ToTTenTranz]

It's somewhat funny that nowadays people expect to play AAA games with ease using over 10 year old hardware. Good luck running Crysis with Pentium 2 and ATI Rage Pro. It used to be a norm that your hardware was obsolete after like two years.

Hardware wasn't obsolete in 2 years. My 9700 Pro was still a perfectly capable GPU after two years. Even though people would upgrade much more often because it didn't cost 1000€.

 

[Keller_TT]

9070 will likely have 8 CUs less, but Frontend (SEs/Primitive Pipes), Backend (ROPs) and Caches should all stay the same, so perf/clock will definitely NOT go down by 12.5% compared to the XT.
And it likely has only 10% less mem bandwidth, too.

I don't think this is Navi 456 and Navi 464 doing a repeat of Vega and AMD can't do either their $1000 6900 XT strategy or amazing 7900 XT strategy. 54 vs 64 is likelier and a good spread.
I expect them to have that performance gap around 17-18% and adjust PPW accordingly.

It's OK if the 9070 is only within 5% faster margin to 4070S. What matters is pricing. $400 for 220W 9070 and +4 GB VRAM would be a clear winner and generate interest when FSR 4 also can hold its own vs DLSS 3.x/4 (leaving aside MFG).

 

[Heartbreaker]

Interesting

It matches 4070 ti super RT. That is a win. Now at what price ??

Black Monkey
4080S
4K 33
2K 77
1080 99

XXXX XT
4K 30
2K 73
1080 97

4070Ti S
4K 28
2K 67
1080 87

谁说跑分强游戏不行的?进来打脸~ - 电脑讨论(新) - Chiphell - 分享与交流用户体验

谁说跑分强游戏不行的?进来打脸~,黑猴子4080S4K 332K 771080 99XXXX XT4K 302K 731080 974070Ti S4K 282K 671080 87,电脑讨论(新),讨论区-生活与技术的讨论 ,Chiphell - 分享与交流用户体验

www-chiphell-com.translate.goog

Didn't one of the leakers say 4080 -5%. That's kind of in the ballpark.

For price, I think $550 is more than fine. This might not be far behind 5070ti, and that's $750.

That's a $200/36% discount. Which is pretty massive. I wouldn't expect less than that. Maybe even $600 would be more appropriate though AMD can do $550 for their own cards, and AIBs can do $600. Even $600 would still be a 25% discount and chances are all the NVidia AIBs will be selling 5070 Ti for $800+ anyway. IIRC all the polls about "How much discount to Buy AMD", that number is usually just 20%. Unlike many "AMD supporters", I don't want to see AMD bleed red ink for my amusement.

$450 for the 9070 likewise positions it well against 5070... With a health 22% discount, and this time, a nice discount against their own top product, which AMD messed up so badly with 7900 XTX/XT.

On top of this pricing maintaining greater than 20%, discount, there all appears to be a shrinking in the feature gap.

If FSR 4 holds up, 20% discount would be enough for me to go AMD instead, as DLSS was the key differentiating for me.

This time AMD appears to have FSR4 with Machine Learning, real Tensor core equivalents, will probably have 4X Fake Frame generation (barf), so the feature gap may be minimal.

 

[SolidQ]

This might not be far behind 5070ti, and that's $750.

Based on 4080 vs 5080, 5070ti might be weaker, than 4080, but that only C2077, we don't have raster perf from RTX 5xxx

 

[Keller_TT]

That's a $200/36% discount. Which is pretty massive. I wouldn't expect less than that. Maybe even $600 would be more appropriate though AMD can do $550 for their own cards, and AIBs can do $600. Even $600 would still be a 25% discount and chances are all the NVidia AIBs will be selling 5070 Ti for $800+ anyway. IIRC all the polls about "How much discount to Buy AMD", that number is usually just 20%. Unlike many "AMD supporters", I don't want to see AMD bleed red ink for my amusement.

Too much hopium for single digit market share. There's no "discount" here. Just greed. You are coming from a long long way back from RDOA 3, and so first AMD must earn the right to be in the discussion.
We haven't seen FSR 4 tested in earnest and only a promising glimpse in 1 game.
Besides, AMD should not price based on Jensen's greed. They'll make more profit and share if they bring real quality and performance to the $400-500/550 range.

 

[GodisanAtheist]

It's somewhat funny that nowadays people expect to play AAA games with ease using over 10 year old hardware. Good luck running Crysis with Pentium 2 and ATI Rage Pro. It used to be a norm that your hardware was obsolete after like two years.

-To be fair, things were moving so quickly back then that two years was like 6 generations of graphics improvements back then.

The bigger issue now is everything keeps hogging more resources but the visual improvements off of good looking 2016 era games just don't seem to justify the squeeze.

 

[SolidQ]

The bigger issue now is everything keeps hogging more resources but the visual improvements off of good looking 2016 era games just don't seem to justify the squeeze.

Looks still Nice

 

[Heartbreaker]

It's somewhat funny that nowadays people expect to play AAA games with ease using over 10 year old hardware. Good luck running Crysis with Pentium 2 and ATI Rage Pro. It used to be a norm that your hardware was obsolete after like two years.

I was happy using my 9700 Pro for 5+ years. Same for my 8800GT which replaced it.

 

[SolidQ]

Same for my 8800GT which replaced it.

Your 8800GT lived 5 years? Mine like 1 year, then changed to HD 4850

 

[Heartbreaker]

Your 8800GT lived 5 years? Mine like 1 year, then changed to HD 4850

I'm never expecting to run latest AAA at maximum resolution with Ultra everything...

 

[GodisanAtheist]

I was happy using my 9700 Pro for 5+ years. Same for my 8800GT which replaced it.

That's wild considering the lack of pixel shader 3.0, which did start becoming mandatory on games pretty quickly in 2007 (I think BioShock was one of them).

I remember being kinda upset that my X800pro didn't support PS3.0 while my friend's 6600 did.

Things definitely slowed down and stabilized a lot more with programmable shaders post 8800/2900 era, but this whole thing is a major digression and I'ma stop talking about before I draw *whispers* you know who's ire.

 

[Keller_TT]

Looks still Nice

I actually think RDR2 is the greatest modern video game ever made. Let's say starting from PS3, Xbox 360 gen of multi-core and exponentially increasing graphical quality.
I bought one just as a collector of a piece of video game masterclass and I've never even attempted to complete it but just fool around.

They could just refine the shaders, textures, and few other techniques, and it'll still look gorgeous for all time, with Zero ray tracing.

 

[Heartbreaker]

That's wild considering the lack of pixel shader 3.0, which did start becoming mandatory on games pretty quickly in 2007 (I think BioShock was one of them).

I remember being kinda upset that my X800pro didn't support PS3.0 while my friend's 6600 did.

Things definitely slowed down and stabilized a lot more with programmable shaders post 8800/2800 era, but this whole thing is a major digression and I'ma stop talking about before I draw *whispers* you know who's ire.

Pretty sure I tried Bioshock but didn't get into it (I have it in my GoG account). Maybe the new enhanced edition or whatever they call it doesn't work on older cards.

I played tons of FO3/NV, and the newest games I remember playing were Witcher 2, and Dragon Age 2 (not as good as DA:O) both are from 2011 and I finished both on my 8800GT.

Note that I mostly play single player RPGs, so not as sensitive to FPS as someone playing multiplayer twitch games.

I can't even think of any 2012-2013 RPGs that I wanted to play, so it got me happily to the 5 year mark...

 

[soresu]

Looks still Nice

Dman that purdy.

 

[gaav87]

Your 8800GT lived 5 years? Mine like 1 year, then changed to HD 4850

Same then 5870 and 7870xt still have them even the 4850 or gt210 sold my 8800gt should have keep it... even have 9700pro and hercules its the size of my 6800xt

 

[reaperrr3]

I don't think this is Navi 456 and Navi 464 doing a repeat of Vega and AMD can't do either their $1000 6900 XT strategy or amazing 7900 XT strategy. 54 vs 64 is likelier and a good spread.

Unless they finally upgraded how fine-grained you can disable WGPs, this is most likely technically impossible.

If previous gens are anything to go by, RDNA can still only disable CUs/WGPs in certain groups, just like GCN, depending on how many active ShaderEngines/-Arrays the chip has.

The full config being 64 CUs suggests 4 SE / 8 SA, which probably still only allows disabling CUs in groups of 8, just like N21. In that case 54 is impossible and 48 would be too low, so 56 is the most likely.
The only alternative is disabling whole SEs, but that would also be only 48 CUs in this case, so...

 

[SolidQ]

8800gt should have keep it

8800 series doesn't long live, alot gpu was died.

 

[Hail The Brain Slug]

Unless they finally upgraded how fine-grained you can disable WGPs, this is most likely technically impossible.

If previous gens are anything to go by, RDNA can still only disable CUs/WGPs in certain groups, just like GCN, depending on how many active ShaderEngines/-Arrays the chip has.

The full config being 64 CUs suggests 4 SE / 8 SA, which probably still only allows disabling CUs in groups of 8, just like N21. In that case 54 is impossible and 48 would be too low, so 56 is the most likely.
The only alternative is disabling whole SEs, but that would also be only 48 CUs in this case, so...

If its 4SE/8SA and they have to disable 4 workgroups, 1 in 1SA per SE, will there be the same penalty in some workloads as the 6800XT where it inherently load balances across both (uneven) SA in an SE causing it to behave as if it only has 48CU? (64CU in the case of the 6800XT)

Or did they fix that with RDNA3? I didn't follow the technical details closely enough to notice.

 

[Keller_TT]

Unless they finally upgraded how fine-grained you can disable WGPs, this is most likely technically impossible.

If previous gens are anything to go by, RDNA can still only disable CUs/WGPs in certain groups, just like GCN, depending on how many active ShaderEngines/-Arrays the chip has.

The full config being 64 CUs suggests 4 SE / 8 SA, which probably still only allows disabling CUs in groups of 8, just like N21. In that case 54 is impossible and 48 would be too low, so 56 is the most likely.
The only alternative is disabling whole SEs, but that would also be only 48 CUs in this case, so...

Didn't they disable just 6 CUs from Navi 32 for 7700 XT for 54 vs 60 CUs.
Also if I'm not mistaken, they kept the same configuration ratios from RDNA 2 to RDNA 3 and disabled 12 CUs from big Navi for 7900 XT and 16 CUs to make the 7900 GRE.
I'm just thinking they'll keep the same CU weightage for the next-gen of 7700 XT class card it was a 35% increase from RDNA 2 already, and they just massively under delivered. And 54 vs 64 for this gen would be a good split in PPW.

 

[Saylick]

All_The_Watts (formerly rogane I think) says 9070XT is on par or slightly faster than the 4070 Ti Super but not quite the 4080. It also hits >3 GHz with power above 3000W.

9070 non-XT is faster than 4070 Super, but slower than 4070 Ti Super. Clocks in the 2 GHz range with power also in the 200s W.

https://twitter.com/x/status/1878490567014777077

 

[reaperrr3]

Didn't they disable just 6 CUs from Navi 32 for 7700 XT for 54 vs 60 CUs.
Also if I'm not mistaken, they kept the same configuration ratios from RDNA 2 to RDNA 3 and disabled 12 CUs from big Navi for 7900 XT and 16 CUs to make the 7900 GRE.
I'm just thinking they'll keep the same CU weightage for the next-gen of 7700 XT class card it was a 35% increase from RDNA 2 already, and they just massively under delivered. And 54 vs 64 for this gen would be a good split in PPW.

N32 had 3 ShaderEngines, so they could only disable CUs in groups of 6.
N31 has 6 SE -> can only disable in 12CU-steps, as long as no SE are disabled.

GRE is different because it disables 1 entire SE, instead of 1 WGP per SE.

The original 70CU-7800XT based on N31 that AMD were considering had 70 CU because it would've been like GRE, except with 1 additional WGP disabled per active SE.