Discussion RDNA4 + CDNA3 Architectures Thread

Page 56 - 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
1,586
5,692
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:

Saylick

Diamond Member
Sep 10, 2012
3,048
6,090
136
Last edited:

SteinFG

Senior member
Dec 29, 2021
385
442
106
GDDR7 will also have 3gb modules so 192 bit can support 18GB from 6 chips and 128 but could support 12GB from 4 chips.
But that's not coming right away. 2GB GDDR6 chips came out two years after 1GB GDDR6. I wouldn't expect 3GB GDDR7 on rtx 50 or rx 8000, probably a gen after that
 

Timorous

Golden Member
Oct 27, 2008
1,514
2,453
136
But that's not coming right away. 2GB GDDR6 chips came out two years after 1GB GDDR6. I wouldn't expect 3GB GDDR7 on rtx 50 or rx 8000, probably a gen after that

Micron should be making them available at launch unless things have changed.

TPU Source

37AjDk40Z3wosMh7.jpg


Mid 2024ish according to this.
 

eek2121

Platinum Member
Aug 2, 2005
2,869
3,842
136
Yea that's the plan.
Yep. AMD would absolutely love to sell you a tiny GPU at insane clocks. Smaller chips mean higher margins. Tiny chip + overbuilt board cost less than multiple large chips + overbuilt board.

…when it works.

Would you care if your GPU had 40CUs at 5ghz or 60CUs at 2.5? Hypothetical example only.
 

moinmoin

Diamond Member
Jun 1, 2017
4,894
7,568
136
Would you care if your GPU had 40CUs at 5ghz or 60CUs at 2.5? Hypothetical example only.
Personally I'd always pick the more efficient product among the same class of performance, so I'd prefer the latter. But DIY desktop users in general seem to be perfectly fine with the former, and the development of CPU and GPU TDPs over the past decade reflects just that.
 

adroc_thurston

Golden Member
Jul 2, 2023
1,653
1,864
96
Personally I'd always pick the more efficient product among the same class of performance, so I'd prefer the latter
That only ever worked back when xtors were getting cheaper which is just no more.
and the development of CPU and GPU TDPs over the past decade reflects just that.
Well that's just dennard scaling dying off.
Server CPUs had even granted power envelope bumps.
Romley Sandy-EP Xeons were like 95W and Turin will be 5 times that in a bit over a decade.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,312
2,798
106
Yep. AMD would absolutely love to sell you a tiny GPU at insane clocks. Smaller chips mean higher margins. Tiny chip + overbuilt board cost less than multiple large chips + overbuilt board.

…when it works.

Would you care if your GPU had 40CUs at 5ghz or 60CUs at 2.5? Hypothetical example only.
I would care because that 40CU at 5GHz would be faster than 60CU at 2.5GHz. ;)

If you fixed It to the same TFlops 40CU at 5GHz vs 80CU at 2.5GHz then from a performance point I wouldn't really care, but as @moinmoin mentioned, the later should be more power efficient, but because It's bigger It would be also more expensive to make.
 

adroc_thurston

Golden Member
Jul 2, 2023
1,653
1,864
96
If you fixed It to the same TFlops 40CU at 5GHz vs 80CU at 2.5GHz then from a performance point I wouldn't really care, but as @moinmoin mentioned, the later should be more power efficient, but because It's bigger It would be also more expensive to make.
It's not quite that.
RDNA2 clocked higher while being more efficient.
 

Mopetar

Diamond Member
Jan 31, 2011
7,784
5,879
136
That s not even enough in respect of a 12 CUs APU that is 178mm2, let alone one that has 32CUs as well, indeed German etailer Mindfactory sales numbers are indicative that the 7600 lack perfs to be competitive, at 40CUs it would have sold much better while at 32 it s buried within a plethorous offering.

In this segment people get either a RX 6600 wich is cheaper and perform not that much worse or RTX 3060/4060/4060TI wich are either price competitive for the weakest or somewhat faster for the others.


The real story is that when you offer a great value product like the 7800XT you actually get a lot of people to buy your cards.

AMD could have definitely had more success with their other products, but they decided to chase extra $$. However, it's hard to blame them when it's probably much more profitable to sell under half as many cards if you can get an extra $50 on each of them.
 
  • Like
Reactions: Tlh97
Aug 4, 2023
94
163
66
N44 1SE/2SA/16WGP 24MB MALL 96-bit GDDR7 9GB
N48 2SE/4SA/32WGP 48MB MALL 192-bit GDDR7 18GB
N44 looks sad, but will be stronger than N33 so it should be a bargain.
N48 on the other hand could be sold for 7700XT prices while beating N32/N21 and giving cutdown N31 a real scare.
 
Last edited:
Aug 4, 2023
94
163
66
Let's get the hypetrain rolling on 3.5GHz+
That is the goal, as clocks are still untapped potential in GPU design.
The traditional way of scaling GPU performance and cost structure, Moore's Law, is dead.
Increasing clocks, dark Si and trimming off all the fat to make a very performance dense part is the winning strategy, on client at least.
Now chiplets will lead to lower costs per equivalent yielded area, but packaging does incur another hit. Still compute demands will require beyond reticle designs to increase compute density, this is the way in DC and halo client. Memory is the really big thing that needs to try to keep up with compute demands, along with interconnects and scale out fabrics.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,312
2,798
106
N44 1SE/2SA/16WGP 24MB MALL 96-bit GDDR7 9GB
N48 2SE/4SA/32WGP 48MB MALL 192-bit GDDR7 18GB
N44 looks sad, but will be stronger than N33 so it should be a bargain.
N48 on the other hand could be sold for 7700XT prices while beating N32/N21 and giving cutdown N31 a real scare.
N44 in my opinion will have 128-bit 8-12GB. 9GB is just too exotic.
I think 32MB IC(MALL) is more realistic than only 24MB, which doesn't even have 50% hitrate at FullHD.
 
Last edited:
  • Like
Reactions: Tlh97

gdansk

Golden Member
Feb 8, 2011
1,903
2,222
136
Is It really just a hypetrain?

7700XT shows very high OC.
View attachment 90028
And that is with +15% power limit. With higher limit, I think you could clock It higher.
The architecture is capable to go >3GHz, but they need to fix power consumption.
The question is If they fixed It in RDNA3.5 and RDNA4 or not.
I don't know if hypetrain is the right word but N48 will be a real flop if it isn't hitting such high clocks.
 

HurleyBird

Platinum Member
Apr 22, 2003
2,669
1,248
136