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

 

[Win2012R2]

Very disappointing the needle is not moving 15-20% past 3 Ghz, perhaps there is no point since the memory is old tech and "infinity cache" on those parts is now tending to zero.

 

[gaav87]

Allegedly, according to the leak, it performs within 5% of a 4080 in games. What those games are? Who knows.
View attachment 113807

Tested in 1080p with 9600x (no pbo) + 5200mhz ram +fsr performance.
🙄no cpu bottleneck at all

 

[Win2012R2]

What those games are? Who knows.

Could be (mild-medium) ray traced stuff, this is probably where the biggest relative improvement (from crap base) in RDNA4 is.

 

[SolidQ]

Could be (mild-medium) ray traced stuff, this is probably where the biggest relative improvement (from crap base) in RDNA4 is.

it's China, so think wukong is there, but wukong pretty heavy NV game.
Alan Wake in RT mode also pretty NV game, Cyber 2077 is different story Raster pro AMD, RT pro NV

 

[Win2012R2]

AW2, Cyber - all way too heavy ray tracing games, I tried to play the latter in RTX mode in 3090 but turned that off, then tried again in 4090 and still I don't think it's the perf wise - visual upgrades were just not there for loss of perf, maybe finally 5090 can run it at acceptable framerates so it will be "ok why not turning it all on", but these are real outliers. Have not tried Wukong yet (all new release gaming is on pause for me until 5090 is out).

Most likely they've tried other stuff - say Indians Jones might be very good candidate, especially with its high memory usage.

 

[Joe NYC]

3.1 would mean ~= 7900XT on compute, but no more. And 20% less bandwidth.

Have there been any leaks on the Navi48 die size?

 

[gdansk]

Have there been any leaks on the Navi48 die size?

Yes but they're all incorrect so far.

 

[gaav87]

9070 xt > 5080
confirmed ? xD

 

[gaav87]

Im just messing found this funny.
For 9070xt > 5080 You just need:
3500 fMax and 380w (+15% power slider) from the "leaked" 330w 9070xt (from chiphell) and 2750mhz on mem oc and you good

 

[itsmydamnation]

AW2, Cyber - all way too heavy ray tracing games, I tried to play the latter in RTX mode in 3090 but turned that off, then tried again in 4090 and still I don't think it's the perf wise - visual upgrades were just not there for loss of perf, maybe finally 5090 can run it at acceptable framerates so it will be "ok why not turning it all on", but these are real outliers. Have not tried Wukong yet (all new release gaming is on pause for me until 5090 is out).

Most likely they've tried other stuff - say Indians Jones might be very good candidate, especially with its high memory usage.

you mean path tracing ? ray tracing in CP2077 , witcher 3 and the the few others i have tried seem fine on 7900xtx , what i have found is having even very basic RT GI/ shadows / reflections , adds a lot ( shadows less so ) at very modest perf hit but i cant tell the difference when turning up the RT quality levels/ number of bounces.

At least to me the first order effect is like 99% of the value in RT. so if NV does have a significate advantage in RT at a given price point i expect dumb shenanigans/ pre-set configurations.

 

[Win2012R2]

9070 xt > 5080
confirmed ? xD

Saw this, but find it very unlikely that 5080 will get double mem bandwidth and boost only to same levels as before, considering that card won't be cheap this is bull for sure

 

[gaav87]

Saw this, but find it very unlikely that 5080 will get double mem bandwidth and boost only to same levels as before, considering that card won't be cheap this is bull for sure

OFC it's bull.
About the double mem b/w.
4070ti -> 4070ti super
+7% cuda's +33% b/w
+9% performance (TPU new perf chart)
So only cuda and b/w mean nothing w/o arch changes and clock speed increase.

 

[RNR_Forte]

The rumor mill is a rumor tornado. Sincerely i even not know what rumor is true if performance is kinda like =~ 4080S raster and rt or not.


PS: the only thing i know is that AMD buyers/fans will be surprised positively.

 

[gdansk]

based on specs it is a cheaper to make 7900gre with more features. slightly faster. but apparently better RT.

don't fall for the overhype.

 

[RNR_Forte]

I apologize to AMD the RDNA4 is all about price to performance.

 

[ToTTenTranz]

Saw this, but find it very unlikely that 5080 will get double mem bandwidth and boost only to same levels as before, considering that card won't be cheap this is bull for sure

Nvidia could be taking advantage of GDDR7's higher bandwidth to decrease the L2 size and produce a smaller chip, i.e. make more money out of the same performance tier.

 

[branch_suggestion]

Nvidia could be taking advantage of GDDR7's higher bandwidth to decrease the L2 size and produce a smaller chip, i.e. make more money out of the same performance tier.

The bandwidth bump is for offline rendering and ML perf beyond anything else.
Gaming it will help, but CPU bottlenecks are becoming an increasing problem unless you crank the fps virus.

 

[Win2012R2]

Nvidia could be taking advantage of GDDR7's higher bandwidth to decrease the L2 size and produce a smaller chip

I reckon cache is only going up with Nvidia, at least for 4080/4090 levels, it's only AMD that cuts "infinity cache", I guess technically it still legit as half of infinity is still infinity. 4080 is clearly being positioned to take place of a 4k card, this will need good cache.

Here is the useful reminder of cache hit rate depending on size and resolution - 128 MB had only 62% hit rate at 4k, and textures ain't getting any less high rest now compared to 2020. They did not show how far it goes, but clearly 256 MB would have have provided awesome hit rates, just going up from 62% to 80% would HALF memory accesses! It was clearly necessary to get cache up to 192 or even 256 MB, then maybe 256 bit memory interface will be ok, but cheapMD took the other view.

 

[Win2012R2]

The bandwidth bump is for

true high refresh rate* 8k gaming, you've heard it on here first...


* using Super Ultra Performance DLSS 5.0

 

[adroc_thurston]

but cheapMD took the other view.

They didn't, N31 (and N36) had an X3D variant planned

 

[Arctic Islands]

Have there been any leaks on the Navi48 die size?

250-300mm^2 for the GCD. Total die area is >400mm^2.

 

[Joe NYC]

250-300mm^2 for the GCD. Total die area is >400mm^2.

But RDNA4 was supposed to be monolithic, for both dies. Isn't even that settled?

 

[marees]

But RDNA4 was supposed to be monolithic, for both dies. Isn't even that settled?

No hope for navi 48 if it is not monolithic

All the bugs of RDNA 3 will then remain

 

[adroc_thurston]

All the bugs of RDNA 3 will then remain

what

But RDNA4 was supposed to be monolithic

it is.

 

[reaperrr3]

I reckon cache is only going up with Nvidia, at least for 4080/4090 levels (...)
4080 [sic: presumably meant 5080] is clearly being positioned to take place of a 4k card, this will need good cache.

No offense, but that's wishful thinking.

The 5080 will rely mostly on GDDR7 to provide the additional bandwidth needed. 256bit@32Gbps is roughly the same VRAM bandwidth as the 4090, and the 4090 has only 72MB L2 active, so 64MB (like on the 4080) is close enough to hit similar performance levels, if the raw performance is there.

GB203 is ~ the same die size as AD103 with the same number of SM on virtually the same process. That makes anything more than 64MB L2 exceptionally unlikely, as 128MB would definitely increase die size substantially.

The only Blackwell that will (probably) get more cache than its predecessor will be GB202, at 128MB.
But that's in part simply because with a 512bit interface, the alternative would've been 64MB (256MB would've probably made it exceed the reticle limit, so not really an option).
Although I wouldn't be surprised if they disabled 25% of it again, since VRAM bandwidth is up by so much that 96MB L2 will be a sufficient increase over the 4090.