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

 

[gaav87]

16WGP per SED, I wonder the SA config.

It is the first GPU in a long time that has >20% perf headroom with pushing power beyond the design power. The thing is broken.

Well RDNA4 does aim to simplify the WGP in clever ways to make it more spammable. That would only make clocks go even higher, rather than increasing per unit throughput.

Not a chance in a million years that a BW SM will be equivalent to an RDNA4/5 WGP.
Stop pretending that you know anything, I grow tired of it

.

Hm did not know that rdna4 aimed to simplify the wgp. Maybe its 8x simds that are 16wide with 16/8 wave slots per simd but each simd can access other simd data. This would be similar to rdna3 dual issue but if they could access the data then it would work better in gaming vs rdna3 (+1-6% dual issue gain). Would be similar to GCN but with 6x lower ocupancy levels so no wasted time/latency on waiting. This would simplify the gpu and allow for that 3ghz+?

rtx4000 toping out at 2600mhz on N4 they are packing 32wide. Maybe going 16wide lowers the voltage leak ?

 

[dhruvdh]

https://twitter.com/x/status/1784561456694046744

Quite sad to see the N40 and N41 canned. Looks like an ambitious design to say the least.

The most interesting part of tweet to me seems to be Navi 36. Why are others not discussing it more? Anyone has any more info?

 

[Mahboi]

The most interesting part of tweet to me seems to be Navi 36. Why are others not discussing it more? Anyone has any more info?

Clearly it was canceled.
Probably after realising that RDNA 3 was what it was. I mean, if they're going to cancel all mid to high end RDNA 4 and it supposedly works well, imagine how quickly they must've canned highest end RDNA 3 after they realised it was an awful power hog.
Not much more to say than they at least planned to respond to the 4090 amount of SMs at some point.

 

[dhruvdh]

The code name suggests development was started after 31. Similar to 40,41 -> 48.

Could it be a RDNA3.5 product that is yet to launch? Because as much as RDNA 3 did not meet expectations, it still has been selling well. Even it 36 did not meet expectations, was there really a need to cancel it? Nvidia seemed to have make it okay to have 600W GPUs.

 

[Ghostsonplanets]

The code name suggests development was started after 31. Similar to 40,41 -> 48.

Could it be a RDNA3.5 product that is yet to launch? Because as much as RDNA 3 did not meet expectations, it still has been selling well. Even it 36 did not meet expectations, was there really a need to cancel it? Nvidia seemed to have make it okay to have 600W GPUs.

Kepler (Or was someone else) already talked about it. It was a Halo product that was cancelled. AMD cancelled a lot of GPUs and APUs and reorganized their efforts.

Also, RDNA 3.5 is for APUs only.

 

[Mahboi]

The code name suggests development was started after 31. Similar to 40,41 -> 48.

Absolutely not?
The codename indicates that it started as one of the x amount of projects in the Navi 30 lineup. They don't just add more stuff over time, they pick a bunch of options in the early development cycle and pursue or kill depending on how it seems to turn out.

Could it be a RDNA3.5 product that is yet to launch? Because as much as RDNA 3 did not meet expectations, it still has been selling well.

Nope. All RDNA 3.5 will be from now on is a Vega replacement. Cheap, small CUs in all the APUs.

Even it 36 did not meet expectations, was there really a need to cancel it? Nvidia seemed to have make it okay to have 600W GPUs.

An 96 CU XTX reaches its top perf at around 700W.
You wanted to see what a 120 CU RDNA 3 could be????

 

[dhruvdh]

No need to get so worked up, it was just much more interesting to me than the 16 GB RDNA 4 products.

 

[marees]

No need to get so worked up, it was just much more interesting to me than the 16 GB RDNA 4 products.

Apparently RGT mentioned it once

(But no sightings in LLVM patches unlike navi 48)

Page 82 - Discussion - RDNA4 + CDNA3 Architectures Thread

Page 82 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[dhruvdh]

If they revive it as W7950 or W8900 or something that would be amazing.

 

[Mahboi]

Not happening.
RDNA 3 is gone. Give it 4 months and it'll be official.

 

[Shmee]

Not happening.
RDNA 3 is gone. Give it 4 months and it'll be official.

What will replace RDNA 3 then? Particularly, what will happen at the high end? Will RDNA 4 be faster?

 

[branch_suggestion]

What will replace RDNA 3 then? Particularly, what will happen at the high end? Will RDNA 4 be faster?

Slower than the 7900XTX so that might live for a bit longer, rest of the stack will go EOL very quickly though.

 

[jpiniero]

Slower than the 7900XTX so that might live for a bit longer, rest of the stack will go EOL very quickly though.

Given how AMD products have gone of late, that's unlikely to happen, haha.

 

[Shmee]

Yeah that is kind of what I am thinking, as long as the 7900XTX is faster than RDNA 4, it will still be a viable option, at least until RDNA 5.

 

[GodisanAtheist]

It's nice to know AMD had planned to go "big" with RDNA3 and just chose not to. Makes the "we could have beat AD102 we just didn't want to" sound less neck-beardy.

It was always a little baffling that AMD did chiplets and then maxed out with a 300mm2 GCD. 400mm2 would have absolutely been competitive with the 4090.

 

[SolidQ]

It's nice to know AMD had planned to go "big" with RDNA3 and just chose not to

Maybe due high power consumption it's canceled

 

[Mahboi]

What will replace RDNA 3 then?

Literally nothing until 2026.

Particularly, what will happen at the high end?

A 240mm² die will become the "high" end.

Will RDNA 4 be faster?

For its size, sure. Hugely so.
It's still a 240mm² die.

Page 80 - Discussion - RDNA4 + CDNA3 Architectures Thread

Page 80 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Considering how much Lisa loves her margin, expect it to be above $500.

 

[Mahboi]

Given how AMD products have gone of late, that's unlikely to happen, haha.

Yeah I expect that since they did everything to run away from RDNA 3, they'll phase it out extremely fast.

 

[Joe NYC]

A 240mm² die will become the "high" end.

For its size, sure. Hugely so.
It's still a 240mm² die.
[
Considering how much Lisa loves her margin, expect it to be above $500.

With lower manufacturing costs than 7800 XT, margin will be higher than 7800 XT even if it is priced at $499.

 

[Mahboi]

With lower manufacturing costs than 7800 XT, margin will be higher than 7800 XT even if it is priced at $499.

Of course it will be. They're doing N33 all over again:
200mm², TSMC6 when TSMC 5 was available, monolithic, small bus, GDDR6
240mm², TSMC4P which shouldn't be under high pressure, monolithic, bigger bus though, GDDR6 when GDDR7 is available, and GDDR6 18Gbps to boot

But it's Lisa's AMD. They care about margins more than anything.
So I expect them to make it for $550 and slowly lower to $450 over 18 months.
It'll be a great GPU for 1440p, it's just that I don't expect anything out of AMD pricing anymore. It's just "make cheap, sell as high as you can". Honestly annoys me.

 

[H T C]

Problem with AMD is that they aren't "as ruthless" as nVidia.

Pricing their cards SLIGHTLY cheaper than nVidia @ same or slightly better performance IS NOT ENOUGH: AMD needs to "aim for the jugular" and either have a similar price for SUBSTANTIALLY HIGHER PERFORMING cards (i mean like 40%+ higher), or MUCH CHEAPER price for similar performance cards (i mean like 40%+ cheaper), EVEN IF they take a monetary loss in the immediate future BECAUSE of it.

That's the only way i see AMD gaining ground on nVidia.


AMD couldn't afford to take this approach before ZEN, but THEY CAN now.

 

[Mahboi]

Problem with AMD is that they aren't "as ruthless" as nVidia.

Pricing their cards SLIGHTLY cheaper than nVidia @ same or slightly better performance IS NOT ENOUGH: AMD needs to "aim for the jugular" and either have a similar price for SUBSTANTIALLY HIGHER PERFORMING cards (i mean like 40%+ higher), or MUCH CHEAPER price for similar performance cards (i mean like 40%+ cheaper), EVEN IF they take a monetary loss in the immediate future BECAUSE of it.

You literally just described what they did in Thermi era.
It didn't work then and will work way worse now.

That's the only way i see AMD gaining ground on nVidia.
AMD couldn't afford to take this approach before ZEN, but THEY CAN now.

AMD needs a compelling product before they price war. You're totally misreading the market.

If you don't have a strong offering, the competition just keeps its prices high and says "so what? We got the best, we ain't lowering it for the peasants". If anything Nvidia even mastered the manipulative practice of spreading FOMO, at this point, selling their cards for an absurdly high price is presented as a "sign of Nvidia quality". If they lowered price, they'd lower the brand, as if they're selling a f*****g luxury bag.

First a stronger card, that puts pressure on Nvidia, keep the prices high until Nvidia caves and has to lower prices because they aren't selling enough anymore.
Then when NV prices are low enough, AMD will have to start lowering. Anything else will just be AMD lowering margins and people will still buy "Nvidia quality" that self-justifies just by being more expensive.

 

[Shmee]

Literally nothing until 2026.

A 240mm² die will become the "high" end.

For its size, sure. Hugely so.
It's still a 240mm² die.

Page 80 - Discussion - RDNA4 + CDNA3 Architectures Thread

Page 80 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Considering how much Lisa loves her margin, expect it to be above $500.

I am a bit confused here, as it seems you may be saying contradicting things, at least to me. You say that nothing will replace RDNA 3 at the high end, but then say that an 240mm die will be the high end. Assuming this new die is RDNA 4, either it is faster and it replaces the 7900XTX, or it is slower and it doesn't. It may be newer, but if it isn't faster, then it isn't the new top end card.

I suppose I should clarify that by high end, I meant 7900XT and above, mainly thinking the XTX. Still, if the 7900XTX remains the fastest AMD card for some time, then it isn't replaced, so long as there is stock, new or used. I suppose they could stop producing them, and supply could start to dry up, but that doesn't make the 7900XTX any slower. Replacing it would seem a very bad move, if there isn't anything faster.

 

[H T C]

You literally just described what they did in Thermi era.

They were pricing their cards 40+% lower @ similar performance, or pricing it similar but with 40%+ more performance, then?

 

[marees]

So I expect them to make it for $550 and slowly lower to $450 over 18 months.

Recently the 7900 GRE was discounted briefly for $510 & the 7800 xt for $450

If navi 32 sells out before RDNA 4 launch, then AMD could go for more aggressive launch price for RDNA 4, I think

Otoh, if Navi 32 sticks around then my best guess is:

N44 = 4060 ti = $300
7700xt = $350
7800xt = $420
N48 cut down
N48 = 7900xt / 4070 ti / 4070 ti super = $550