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

 

[sl0519]

7700XT vs 7600XT also had 44.4% difference in 4K, yet there was nothing between them.

60 class needs to be 192-bit bus and 12G of ram by today’s standard. It feels like a total ripoff leaving that much of performance gap, similar to how Nvidia “treats us” gamers.

 

[Josh128]

$299 tops?

MRSP
Radeon RX 7600 $269
Radeon RX 7600 XT $329
Radeon RX 7700 XT $449
Radeon RX 7800 XT $499

Radeon RX 9700 $549
Radeon RX 9700 XT $599
Guys, please be more realistic about AMD.

Die size may be smaller than 7600/7600XT, but the bigger reason is, how well did those crappy SKUs actually sell?? If they want to make it compelling, they need to find a way.

 

[TESKATLIPOKA]

no.
That would be 8+4 where a third of your memory has half the membw.

4+2 in case of 16Gbit chips.
Yeah, anything over 8GB would have only 1/2 of BW in case of a 128-bit bus.
In theory It is possible, but you get a pretty severe performance hit.

 

[jpiniero]

I could see nVidia doing a 5060 Ti Super with 12 GB only (with 3 GB chips)

Maybe AMD will do the same if N44/48 supports GDDR7.

 

[Abwx]

If they didnt implement 40 CUs for the 9060XT they ll regret it.

 

[jpiniero]

If they didnt implement 40 CUs for the 9060XT they ll regret it.

It is most likely the same as N33 (32)

Perf should be close... well close enough... to the 5060 Ti.

 

[marees]

If they didnt implement 40 CUs for the 9060XT they ll regret it.

Not anytime soon.

Probably after 9 months you might the 48 CU part

 

[SolidQ]

If they want to make it compelling, they need to find a way.

they missed it, Best way for them would be at least 10GB for 250-270$ with ~7700XT perf. That where is market share, but none of these Sku's is support 160/10Gb

 

[TESKATLIPOKA]

It wont. Now that Adroc (assuming hes correct) has said half of everything and less-- I'd be willing to bet the best we'll see is 6700XT +10% raster, a bit better than that RT. <180mm die, 3.1 GHz, fast and cheap. Now the insistence of slapping 16GB on there is a mistake, IMO. That could push price north of $300 for absolutely zero benefit to the gamer.

Apparently they wont. But just like the 7600XT 16GB, its going to be wasted.

It offers a performance increase.
7600XT 16GB vs 7600 8GB
FullHD: +10.3%
QHD: +11.4%
4K: +33%
Ok, 12GB would most likely offer the same performance uplift at least today, but I would rather pay a few bucks more for the extra 4GB Vram and 12GB is not possible on a 128-bit bus, until 24Gbit memory chips are available.

 

[GodisanAtheist]

no.
That would be 8+4 where a third of your memory has half the membw.

- GTX 970 3.5+.5 memory scandal part deux, let's gooooooookk

 

[TESKATLIPOKA]

If they didnt implement 40 CUs for the 9060XT they ll regret it.

And what would they gain by putting 40CUs inside? That chip would be bigger, but performance won't increase by 25%(40/32) because of limited BW.
Yeah, they can lower the clocks by 20% to gain efficiency, so instead of 3GHz only 2.4GHz, but It will increase the manufacturing cost.

edit: If It used GDDR7 then with 28-30gbps(+55.6%) memory you could feed even a 48CU GPU. Don't understand why Nvidia didn't increase the SM count for the GB206 If they increased BW by such a massive amount.

 

[TESKATLIPOKA]

- GTX 970 3.5+.5 memory scandal part deux, let's gooooooookk

That was not because of clamshell to be precise and the penalty wouldn't be as severe, but It would affect a bigger part of the whole memory.
So no 12GB on 128-bit until 24gbit memory is out.

 

[lixlax]

Maybe for bottom perf? With 9060 AMD need 7700XT and 7800XT perf, for below perf should be 9050
5060ti should be like ~4070 perf, 5060 ~4060ti

 

[Tuna-Fish]

edit: If It used GDDR7 then with 28-30gbps(+55.6%) memory you could feed even a 48CU GPU. Don't understand why Nvidia didn't increase the SM count for the GB206 If they increased BW by such a massive amount.

The 60-series probably won't use GDDR7?

NV has used cheaper ram for the 60-class and below since the 30-series.

 

[Josh128]

It offers a performance increase.
7600XT 16GB vs 7600 8GB
FullHD: +10.3%
QHD: +11.4%
4K: +33%
Ok, 12GB would most likely offer the same performance uplift at least today, but I would rather pay a few bucks more for the extra 4GB Vram and 12GB is not possible on a 128-bit bus, until 24Gbit memory chips are available.

The full HD and QHD uplifts are almost solely due to 10% higher game clock and 5% higher boost clock.

 

[Tup3x]

- GTX 970 3.5+.5 memory scandal part deux, let's gooooooookk

GTX 550 Ti would be better example (although since it's mixed density stuff, I'm not sure...). I actually have one as backup card (which is somewhat pointless now that I also have GTX 660 Ti as backup card).

 

[Ghostsonplanets]

Lack of Media Engine on N44 wouldn't be a problem this time as basically most CPUs sold currently have their own Media Engine/iGPU. It was more dramatic back in N24 due to Ryzen 5000 and below (non-G) lacking iGPU.

 

[Timorous]

how do I 12GB on a 128b part.

You cut the bus to 96bit but not on the 9060XT. On the cheaper $249 part though I think that is a better and more marketable compromise than an 8GB 128bit card.

 

[Timorous]

Sorry, but why do you think 9600XT 16GB would cost only $329?
Both RX 9070(XT) have $100 higher MRSP than 7700XT or 7800XT at launch, so there is no way AMD would price It at the same price as 7600XT.


Exactly. You get higher performance both in raster and even more in RT, better features etc. There is no way AMD won't increase the price.

Because based on what we know the 9060XT die will be tiny so the thing will be dirt cheap. It would offer similar margin to the 9070XT in a higher volume product.

As for the stack it should be simple

9060XT - 16GB 20gbps ram $330
9060 - 12GB 20gbps ram, cut the bus to 96bit. $250
9050 - rebrand N33 and lower clocks to hit a sub $200 price point.

Any 8GB card for more than $200 will get absolutely shredded in reviews so no point in doing it. A 96bit 12GB card that manages 6700XT tier performance will be far better received than a 128but 8GB card that can sit in the gap between the 6700XT and 7700XT but then suffers from horrid performance/IQ in some of the latest titles due to Vram limits.

 

[jpiniero]

Any 8GB card for more than $200 will get absolutely shredded in reviews so no point in doing it.

Which is why you do both 8 and 16... let reviewers bash the 8 GB model and watch people buy it anyway because of the lower price.

 

[Timorous]

Which is why you do both 8 and 16... let reviewers bash the 8 GB model and watch people buy it anyway because of the lower price.

Why waste supply of chips building 8GB variants that will be slower than Strix Halo in a lot of titles?

 

[jpiniero]

Why waste supply of chips building 8GB variants that will be slower than Strix Halo in a lot of titles?

Strix Halo is not really a comparable product.

 

[Timorous]

Strix Halo is not really a comparable product.

Who said it was? A new 60 class dGPU losing to an iGPU is just a sad state of affairs and should not happen, especially when it boils down to a Vram shortfall.

 

[DaaQ]

Difference between models

90% for RTX 5090, maybe RDNA5 will be closer, but not faster

Hey at least as far as we know, the AMD cards don't have variable ROPS, right?

If they can sell a 9070XT for $600, it shouldnt be. $299 absolute tops. But hey, if they want to lose more market share, they can go right ahead.

Here is another example of 30 dollars one way or the other somehow makes the card extra special or extra craptastic. But does it have the variable ROPS feature?

 

[basix]

I don't care if AMD releases another 8GB model. They shall do it if they see benefit in it.
But besides of that and the 9060XT 16GByte, there should be a 96bit / 12GByte model. That would be the Go-To card if you cannot afford the 9060XT 16GB. A few percent performance drop due to 96bit instead of 128bit is acceptable and far less noticable compared to an overfilled VRAM buffer.

- 9060XT, 16GB, 32CU, 349$
- 9060, 12GB, 28 CU, 289$