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

 

[TESKATLIPOKA]

Oh gawd 4 more CU.
Groundbreaking.

It's infinitely more than big fat ZERO increase with N44 vs N33 in CU count.

 

[moonbogg]

It's a problem for our wallets, not the AIBs. And until they stop selling out immediately, the prices will get higher yet. Just as they did last time the market went derp. You don't have to like it, support it, or buy it. Legions of FOMOs and scalpers ensure you and I are swept up in their idiocy anyways.

Me reading all of the magic 8 balls around here proclaim that if AMD goofed on pricing this time they are cooked

I was one of those who thought that $50.00 would make or break this launch, lol. I should have known better. I should have known they'd fly off the shelves at $8-900 like they did and continue to do. I should have known that desperate gamers would gladly kick their mother in the gut to steal her wallet to pay a scalper $1500 on ebay for a 9070XT. Happens all the time. I should have known that like last time, gamers will sell their children for GPU money. They always do that. These people are sick.

 

[SolidQ]

Interesting review

https://www.pcmrace.com/2025/03/10/rx-9070-xt-aorus-elite-review-analisis/

Some results are good

 

[Timorous]

N48 has more CU than N32, but N44 didn't increase It. So N44 would need additional frequency to compensate.

I mean given the node differences that seems pretty trivial to overcome. Also N33 has a 128KB VREG vs N32/N31 192KB VREG so no idea how that plays a part in performance but the difference is there. It is not like CUs are the only difference.

Point is you can get into the nitty gritty if you want but it seems like massively overkill for working out the rough performance level a 9060XT might have. It should go without saying the error bars are quite large but I guess I will say it explicitly. 7700XT +/- 15% is probably the error bars you are looking at which is a range from 4060Ti at the lower end and 6800XT at the upper end. It will depend on final clocks, the final ram spec, tdp choices and so on. That gives maybe a $300 to $380 pricing range IMO. If it manages towards the upper end of that range then a cut 9060 could also be equipped with 16GB of VRAM because why not if the performance is enough to drive a $300-320 price tag. On the lower end of that range though and you get into the idea of a lower VRAM 9060 in which case I think 12GB + 96bit has fewer compromises than an 8GB 128bit part.

N48 costs more than N32 by $100, that's in my opinion more than enough proof that N44 won't cost the same as N33. My bet is $50 more.

N48 is a more competitive product than N32 so can command a higher price.

N44 will be in the 7700XT class +/- a step. The 7700XT can be had for around £380 in the UK and the 7600XT is £300. That means the max price for a 9060XT is probably around £320 if you want a 20% perf/$ advantage. At current exchange rates that would be around $340 USD when you take off VAT from the UK price.

If you think it will be $380 then the only way I can see that being the case is if performance is towards the upper bound of what I said which is 6800XT performance and that seems optimistic to me.

N33 was also dirt cheap to make. Cheaper than N44 to be honest and we all know what the price was.

Selling price does not always have a lot to do with BOM cost. N33 was also cheaper than N23 was but has sold for more than N23 parts did despite AMD having the option to keep margins steady and sell at a lower price.

I am sure N44 will probably cost more than N33 just through wafer pricing it will also be cheaper than N32 was so there is room for them to sell N44 for less than the 7700XT and still keep the same margins or slightly better than on N32.

 

[Josh128]

Nvidia engineer explains whats happening with 5090 and the rest of the series. This NEVER gets old.

 

[Gideon]

Well, at least Frank Azor is promising better stock also in Europe:

https://twitter.com/x/status/1899169007497871761

 

[blackangus]

View attachment 119396
love this comment lol

Yeah where Nvidia did it more and with less actual cards available at MSRP, in some cases none.
But yeah AMD is the bad guy. Except they delivered 10's of thousands of cards at MSRP....

This person seems to have a crazy lack of grasp on reality.

 

[poke01]

Yeah where Nvidia did it more and with less actual cards available at MSRP, in some cases none.
But yeah AMD is the bad guy. Except they delivered 10's of thousands of cards at MSRP....

This person seems to have a crazy lack of grasp on reality.

In the US, specifically Micro centre. AMD did a brilliant job, I don’t deny that, it’s just they did a brilliant job in one country.

Asia, Europe and SA barely got any and definitely not 10s of thousands in these regions. This is why I say to get rid of current AMD marketing, sometimes it’s better to keep quiet.

 

[blackangus]

In the US, specifically Micro centre. AMD did a brilliant job, I don’t deny that, it’s just they did a brilliant job in one country.

Asia, Europe and SA barely got any.

How does that relate to the point of my comment, in all cases NVidia did worse so that persons post is completely misplaced and delusional.

 

[poke01]

There also the point of AMD not producing enough GPU dies. Also the retailers and distributors increasing prices which AMD has no control over. As well crazy demand, it’s a lost world.

 

[inquiss]

There also the point of AMD not producing enough GPU dies. Also the retailers and distributors increasing prices which AMD has no control over. As well crazy demand, it’s a lost world.

This again. How could the minor player produce enough GPUs to satisfy demand when the market leader that usually produces 80/85% of volume is absent?

 

[blackangus]

There also the point of AMD not producing enough GPU dies.

This is a fallacy as well. They produced exactly as many as their forecasting called for. The reasoning of - NVidia produced a poor and pretty much un-available product, so AMD should have (Magically) filled that gap - Is a mis-understanding of manufacturing, inventory risk management, and market dynamics at best; a completely bogus intentional misdirection at worse.

 

[adroc_thurston]

It's infinitely more than big fat ZERO increase with N44 vs N33 in CU count.

does not make a difference.

In the US, specifically Micro centre. AMD did a brilliant job, I don’t deny that, it’s just they did a brilliant job in one country.

Asia, Europe and SA barely got any and definitely not 10s of thousands in these regions. This is why I say to get rid of current AMD marketing, sometimes it’s better to keep quiet.

AMD can only do so much; their supply chain is not configured to assume NV does not exist.

NVidia produced a poor and pretty much un-available product, so AMD should have (Magically) filled that gap - Is a mis-understanding of manufacturing, inventory risk management, and market dynamics at best; a completely bogus intentional misdirection at worse.

ding-ding-ding.
Hijacking the 5070 launch window was a genius move and it sure as hell worked for their optics, but they physically can't serve all the market demand available.

 

[gdansk]

There also the point of AMD not producing enough GPU dies. Also the retailers and distributors increasing prices which AMD has no control over. As well crazy demand, it’s a lost world.

How were they supposed to know Nvidia simply wouldn't ship? There was no way for AMD to know until it happened.

 

[HurleyBird]

ding-ding-ding.
Hijacking the 5070 launch window was a genius move and it sure as hell worked for their optics, but they physically can't serve all the market demand available.

The right answer.

Yet unless AMD is able to satisfy the market with a sufficiently large quantity of units at (or at the very least near) MSRP over the coming months, then the optics win won't go as far as it could have, and if/when in the future AMD finds itself in a similar situation reviewers (who are less deferential to AMD than Nvidia generally) will be unlikely to give AMD the benefit of the doubt a second time.

 

[adroc_thurston]

and if/when in the future AMD finds itself in a similar situation reviewers (who are less deferential to AMD than Nvidia generally) will be unlikely to give AMD the benefit of the doubt a second time.

well they shouldn't
Just build a halo.
DO IT OR NO BALLS.

 

[HurleyBird]

well they shouldn't
Just build a halo.
DO IT OR NO BALLS.

 

[HuesToo4]

* contingent on having good ip/ppa (which RDNA currently does)

 

[Kepler_L2]

Someone email this to Lisa Su

 

[adroc_thurston]

* contingent on having good ip/ppa (which RDNA currently does)

Vega was the last time AMD had bad PPA.

 

[HurleyBird]

* contingent on having good ip/ppa (which RDNA currently does)

Nah. You aim for the crown always and never stop, no matter what hand you have to play with or what kind of architectural deficit you're suffering.

If Nvidia had cancelled GF100 that generation would have resulted in something like a straight duopoly.

If Nvidia had abandoned the high end with the 200/Tesla series while AMD had went for the crown with 4000/Terrascale, then in one generation AMD would captured most of the market.

If Nvidia had abandoned the high end with GeforceFX as opposed to pulling out every possible stop to make a claim for the crown, Lord only knows how much damage that would have done to them. A lot though.

What killed Radeon as an actual competitor is the positive feedback loop of MBA-brained midwittery that began as a response to the failure of R600. In the hypothetical universe where R600 hadn't failed, then some future failure would have kicked off the self destructive cycle instead.

If AMD had bought Nvidia instead of ATI, as was the original plan, then given time they would have done the exact same thing to Nvidia (assuming Jensen is out of the picture).

 

[adroc_thurston]

What killed Radeon as an actual competitor is the positive feedback loop of MBA-brained midwittery that began as a response to the failure of R600. In the hypothetical universe where R600 hadn't failed, then some future failure would have kicked off the self destructive cycle instead.

Pussyfooting with undersized configs is generally *the* ATi gimmick since R420.
It's kinda adorable how much of that ATi DNA is present still.

 

[Timorous]

Nah. You aim for the crown always and never stop, no matter what hand you have to play with or what kind of architectural deficit you're suffering.

If Nvidia had cancelled GF100 that generation would have resulted in something like a straight duopoly.

If Nvidia had abandoned the high end with the 200/Tesla series while AMD had went for the crown with 4000/Terrascale, then in one generation AMD would captured most of the market.

If Nvidia had abandoned the high end with GeforceFX as opposed to pulling out every possible stop to make a claim for the crown, Lord only knows how much damage that would have done to them. A lot though.

What killed Radeon as an actual competitor is the positive feedback loop of MBA-brained midwittery that began as a response to the failure of R600. In the hypothetical universe where R600 hadn't failed, then some future failure would have kicked off the self destructive cycle instead.

If AMD had bought Nvidia instead of ATI, as was the original plan, then given time they would have done the exact same thing to Nvidia (assuming Jensen is out of the picture).

This.

If AMD had bothered to aim for the top they would have won 5 gens on the trot with 500mm ish designs.

RV770 was ~250mm in the 4870 , 500mm on the same design would have been a 1600 shader monster and given us close to 5870 performance way earlier, it would have crushed GT200 entirely.

Cypress was ~330mm in the 5870, a 500mm design would have been a 2400 shader monster and would have handily bested the GTX 480 and probably gone toe to toe with the GTX 580.

Caymen was also ~330mm so a 500mm design would also have been a 2400 or whatever shader monster that would have been faster than the GTX 580.

Tahiti in the 7970 was also around the 330mm mark so again a 500mm design would have been faster than Hawaii was 2 years earlier, that would have been faster than the GTX 680 and it would have been faster than the GTX 780Ti and the GTX Titan. It probably would have forced NV to release a GK100 based 680 rather than allowing them to go with GK104 and then the refresh GK110 in the 780 would have been a very minor bump.

Hawaii was closer to 450mm so there was room for a 56CU design rather than the 44 they went with. Give it 7gbps ram rather than the 5gbps they did give it and you also have enough bandwidth without needing HBM. That gives you pretty close to the cut down Fury performance years earlier. It would have been faster than the 980 but behind the 980Ti although it probably would have released around the time of the 780Ti so would have been a clear step ahead.

Maybe the extra income from having top tier parts and the extra experience with larger designs would have meant Fiji was actually better than it ended up being.

I know we complain about NV shrinkflation of putting smaller GPUs further up the stack but ATi / AMD actually started that trend and allowed NV to use 330mm dies in the x80 range instead of their older designs that used 500mm designs in that segment.

EDIT: I credit these decisions with where we are today in the GPU market, it allowed NV to become the defacto performance champion and grow the mindshare that comes with it and then they started to do what they did with CUDA and create a walled garden, g-sync, Ray Tracing, DLSS, NVenc etc etc are all just features that allow them to have none hardware reasons to buy their product, it gives them an amount of protection should they misfire which they clearly are with Blackwell but give it 6 months when the supply issues are behind us and I don't think the successful 9000 series launch will make that much difference, primarily because there is no way AMD are going to shift away from their product allocations so they won't try and fill in the gap in the market with extra product.

 

[adroc_thurston]

Hawaii was closer to 450mm so there was room for a 56CU design rather than the 44 they went with. Give it 7gbps ram rather than the 5gbps they did give it and you also have enough bandwidth without needing HBM

See, that's the neat part.
Hawaii was supposed to be HBM until the last minute.

 

[HurleyBird]

Pussyfooting with undersized configs is generally *the* ATi gimmick since R420.
It's kinda adorable how much of that ATi DNA is present still.

Huh? R420 and NV40 were more or less the same size.