Discussion RDNA4 + CDNA3 Architectures Thread

Page 92 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DisEnchantment

Golden Member
Mar 3, 2017
1,659
6,101
136
1655034287489.png
1655034259690.png

1655034485504.png

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
Or Phoronix

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 :grimacing:

This is nuts, MI100/200/300 cadence is impressive.

1655034362046.png

Previous thread on CDNA2 and RDNA3 here

 
Last edited:

jpiniero

Lifer
Oct 1, 2010
14,806
5,431
136
I think low-end Blackwell could be also good depending on specs and price. I was thinking It could use only 96-bit, but I left It only for the cutdown version and only 24gbit modules will be used.

I wouldn't be surprised if the low end Blackwell uses GDDR6 or 6X.

One thing I am wondering about is how far will Corpos go to compromise their products to push AI AI AI. It won't be an issue for NV's high end, because they will just charge enough more... but at the lower end I am thinking the die sizes will be flat at best (assuming N4P)... so I could easily see GB205/6/7 raster game performance being little to none better over the current products.

I do think that N48 will use the x800 series, maybe even x900 just to have something there.
 

Mahboi

Senior member
Apr 4, 2024
741
1,313
96
One thing I am wondering about is how far will Corpos go to compromise their products to push AI AI AI.
Depends on the clients as usual. Uzzi already said that AMD removed the SLC on Strix Point because muh NPUUUUUUUUUUUUUUU from MS.
It won't be an issue for NV's high end, because they will just charge enough more... but at the lower end I am thinking the die sizes will be flat at best (assuming N4P)... so I could easily see GB205/6/7 raster game performance being little to none better over the current products.
Honestly I don't think NV cares for the low end.
 

Mahboi

Senior member
Apr 4, 2024
741
1,313
96
Letting MS determine your product design seems a fools errand.
Preaching to the choir.
But for a ton of huge corpos, a lot of policy isn't decided based on what is smart but on where the money flow will be. No point in finding 10g of gold in a a 10m^3/mn stream if you have the facilities to search for 20g of gold in a 100m^3/mn stream.
AI sure is proving that these days.
 

Mahboi

Senior member
Apr 4, 2024
741
1,313
96
Crazy, and a unique SE:WGP:CU config as well.
Assuming a 50% leap in SMs from Nvidia again (3090 Ti: 80SMs, 4090: 120), we'll see them throw some 160-180 SM thing for Blackwell.
But they were already reaching into 450W+ for the 4090 on N4. Unlikely that they'd make yet another leap of the same size on the same node.

No doubt the 210 CU monster would've wiped the floor with it.
Time to wait for RDNA 5 all over again...