Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

Ajay · Oct 11, 2023

Tigerick said:
There is a wide gap between $500 to $900 without any GPU from AMD, I think I just found out how AMD going to fill it with N52, yes that's mean AMD going to price upcoming 8800XT with 16GB to $699 if my calculation is correct.

I guess I missed something, what is this 8800XT?

Tigerick · Oct 11, 2023

Ajay said:
I guess I missed something, what is this 8800XT?

Replacement of 7800XT, wild speculation of mine

Tigerick · Oct 11, 2023

Ajay said:
He also said N43 wasn’t the correct ID for the die. N31 GCD is 350mm^2

It seems there is another number in play, either N44 or N48???

soresu · Oct 11, 2023

SolidQ said:
Interesting news. Maybe FSR going finally to AI?

https://www.amd.com/en/newsroom/press-releases/2023-10-10-amd-to-acquire-open-source-ai-software-expert-nod-.html

This is more likely AMD making good on their recent promise to double down on software investment, especially in regards to ROCm and AI.

soresu · Oct 11, 2023

Tigerick said:
It is business decision, my friend. AMD is expecting N32 & N33 to have shelf life of two years. Especially with N32 is only available for desktop variant....

Anyhow, below is my speculation and price estimates of RDNA4 & RDNA5 lineups for reference:

SRP 2023 N5/N6 2024 N4P 2025/2026 N3E
$999 7900XTX 24GB 384-bit GDDR6 96 WGP N51 24GB 384-bit GDDR7 210 WGP
$899 7900XT 20GB 320-bit GDDR6 84 WGP N51L 20GB 320-bit GDDR7 ?
$699 N52 16GB 256-bit GDDR7 140 WGP
$599 N52L ? ?
$499 7800XT 16GB 256-bit GDDR6 60 WGP N53 12GB 128-bit GDDR7 70 WGP
$449 7700XT 12GB 192-bit GDDR6 54 WGP
$399 ? ?
$299 ? ?
$269 7600 8GB 128-bit GDDR6 32 WGP

1x WGP = 2x CU + extra stuff (cache?)

The N3x WGP numbers in your table are twice the actual figures.

Tigerick · Oct 11, 2023

soresu said:
1x WGP = 2x CU + extra stuff (cache?)

The N3x WGP numbers in your table are twice the actual figures.

My bad, changing now

Ajay · Oct 11, 2023

soresu said:
This is more likely AMD making good on their recent promise to double down on software investment, especially in regards to ROCm and AI.

Good, they really have to. They are behind NV on hardware and don’t have a sufficiently deep software stack to complete with NV.

adroc_thurston · Oct 11, 2023

Ajay said:
They are behind NV on hardware

No.

Ajay said:
and don’t have a sufficiently deep software stack to complete with NV.

It's alright, just not as simple/sleek/mature as what NV offers, but they're making solid progress anyway.

TESKATLIPOKA · Oct 12, 2023

Tigerick said:
It is business decision, my friend. AMD is expecting N32 & N33 to have shelf life of two years. Especially with N32 is only available for desktop variant....

You still didn't provide any proof that N32 will be limited only to desktop, especially considering you said RDNA4 is not replacing N32.
Even If AMD really kept both N32 and N33 there is no need to slot 2 chips between them, one is more than enough.

Ajay · Oct 12, 2023

adroc_thurston said:
No.

It's alright, just not as simple/sleek/mature as what NV offers, but they're making solid progress anyway.

Seriously? Are you high? Nvidia' product stack, over the next ~year, dwarfs AMD's and their software stack is pretty much untested compared to NVs. NV also has lots of HPC server bits , like high speed fabrics/switches that they can push along with their offering - one stop shopping tends to be a good selling point. And, I guess lastly, NV is already selling a shed load of AI/ML/DNN products.

The best thing AMD has going for it is the ability to bundle or integrate high performance x86 ZEN cores within their offerings.

adroc_thurston · Oct 12, 2023

Ajay said:
and their software stack is pretty much untested compared to NVs

Dawg Frontier exists for a reason.
ROCm is the only viable option outside of CUDA for a good reason.

Ajay said:
NV also has lots of HPC server bits

Every relevant HPC deployment is Cray or Atos.
Some Lenovo.
None are DGX.

Ajay said:
like high speed fabrics/switches that they can push along with their offering

Hyperscale hates Infiniband and the roadmap for that exploded anyway.

Ajay said:
one stop shopping tends to be a good selling point

Yes definitely especially when selling to CSPs who do everything not core processing Si by themselves.

Ajay said:
The best thing AMD has going for it is the ability to bundle or integrate high performance x86 ZEN cores within their offerings.

No they have amazing GPU IP tied to best in the industry packaging engineering.

TESKATLIPOKA · Oct 12, 2023

adroc_thurston said:
No they have amazing GPU IP tied to best in the industry packaging engineering.

GPU IP could be excellent, but If they can't make a great product out of It then It's kinda pointless.
RDNA3 was not a good showing, hopefully we will see some frequency improvement with RDNA3.5.
If not, then It's kinda worrying.

moinmoin · Oct 12, 2023

Ajay said:
The best thing AMD has going for it is

...actually having very decent open source solutions where it matters. Nvidia is still a complete non-starter in areas where this matters.

Heartbreaker · Oct 12, 2023

moinmoin said:
...actually having very decent open source solutions where it matters. Nvidia is still a complete non-starter in areas where this matters.

Where is that?

Ajay · Oct 12, 2023

moinmoin said:
...actually having very decent open source solutions where it matters. Nvidia is still a complete non-starter in areas where this matters.

Uhm, they are selling out of every AI/ML GPU that they have. Nvidia workstation GPUs and RTX 4090s are making their way into those who are trying to get some development done while they wait for H100/H200s to be available. There's not enough packaging capacity to meet demand. I think it's likely that gamers will have a tough time getting GB202 GPUs (maybe NV won't even sell them in the consumer market, since AMD dropped the ball) Their closed source solution really isn't an impediment to developers who just want to get AI integrated into their business processes and software products. The Open source projects are behind the eight ball. People seem to forget all the ground work that NVidia has done getting Cuda and then ML/LLM into universities. They've put in the work and are reaping the rewards (look at their stock price).

AMD is developing good products and improving, but it seems they are around two years behind Nvidia - If AMD can executed near perfectly. Adroc has said they AMD is pouring allot of money into these efforts. I wish them luck. As in all things, competition is good.

moinmoin · Oct 12, 2023

Heartbreaker said:
Where is that?

All the government supercomputer contracts. Incidentally also the reason AMD is where it is today, thanks to DARPA's several Forward sponsoring over the past decade.

Ajay said:
Uhm, they are selling out of every AI/ML GPU that they have. Nvidia workstation GPUs and RTX 4090s are making their way into those who are trying to get some development done while they wait for H100/H200s to be available. There's not enough packaging capacity to meet demand. I think it's likely that gamers will have a tough time getting GB202 GPUs (maybe NV won't even sell them in the consumer market, since AMD dropped the ball) Their closed source solution really isn't an impediment to developers who just want to get AI integrated into their business processes and software products. The Open source projects are behind the eight ball. People seem to forget all the ground work that NVidia has done getting Cuda and then ML/LLM into universities. They've put in the work and are reaping the rewards (look at their stock price).

AMD is developing good products and improving, but it seems they are around two years behind Nvidia - If AMD can executed near perfectly. Adroc has said they AMD is pouring allot of money into these efforts. I wish them luck. As in all things, competition is good.

Not disagreeing with any of that. What to keep in mind though is that Nvidia mostly serves hype markets they create themselves. Nvidia's lock in ensures that that ecosystem can only move into the direction Nvidia wants (which is obviously great business for Nvidia). So any serious enterprise looking to do guaranteed sustainable business in that area has to also look at other options.

Heartbreaker · Oct 12, 2023

moinmoin said:
All the government supercomputer contracts. Incidentally also the reason AMD is where it is today, thanks to DARPA's several Forward sponsoring over the past decade.

Argonne national labs is government funded, and they NVidia GPU based in their supercomputer:

Polaris | Argonne Leadership Computing Facility

www.alcf.anl.gov

Polaris System Specs
Peak Performance
34 petaflops (44 petaflops of Tensor Core FP64 performance)
NVIDIA GPU
A100

So is: Livermore Computing Sierra:

Sierra | HPC @ LLNL

hpc.llnl.gov

GPU Architecture
NVIDIA V100 (Volta)
Total GPUs
17,280

So is, Oak Ridge National Laboratory, Summit:

Summit

The OLCF was established at Oak Ridge National Laboratory in 2004 with the mission of standing up a supercomputer 100 times more powerful than the leading systems of the day.

www.olcf.ornl.gov

GPUs: 27,648 NVIDIA Volta V100s (6/node)

So it appears most government Supercomputer use NVidia GPUs. Where is this supposed lockout?

Ajay · Oct 12, 2023

moinmoin said:
So any serious enterprise looking to do guaranteed sustainable business in that area has to also look at other options.

I don't get this. Nvidia isn't going anywhere. So it is sustainable, because Nvidia has established itself as a reliable partner. The risk to Nvidia, is as more competitors arrive on the scene, that would threaten their 'walled garden' - if those competitors opt for open source alternatives.

adroc_thurston · Oct 12, 2023

Heartbreaker said:
Argonne national labs is government funded, and they NVidia GPU based in their supercomputer:

Polaris | Argonne Leadership Computing Facility

www.alcf.anl.gov

So is: Livermore Computing Sierra:

Sierra | HPC @ LLNL

hpc.llnl.gov

So is, Oak Ridge National Laboratory, Summit:

Summit

The OLCF was established at Oak Ridge National Laboratory in 2004 with the mission of standing up a supercomputer 100 times more powerful than the leading systems of the day.

www.olcf.ornl.gov

GPUs: 27,648 NVIDIA Volta V100s (6/node)

So it appears most government Supercomputer use NVidia GPUs. Where is this supposed lockout?

Those are years old systems.
Now you look at El Capitan and weep.

Heartbreaker · Oct 12, 2023

adroc_thurston said:
Those are years old systems.
Now you look at El Capitan and weep.

Recent AMD Supercomputer wins, is no indication of some kind of government policy blocking NVidia.

Tigerick · Oct 12, 2023

TESKATLIPOKA said:
You still didn't provide any proof that N32 will be limited only to desktop, especially considering you said RDNA4 is not replacing N32.
Even If AMD really kept both N32 and N33 there is no need to slot 2 chips between them, one is more than enough.

Ho, let's go point by point, shall we:

We know there are 2 monolithic dies made by N4P process.
These two dies are positioned like old Polaris, in case people are not familiar of it, have a look of old AT article here. Basically, they are die shrink of current RDNA3 with better power efficiency or lower price or both.
N43/N44/N48 are leak code-names but I don't see the point of N48, do you? N43 actually make sense cause the number is sitting below N32, maybe N48 is like damage control code-name.
I hope nobody is buying GDDR7 BS cause with 192-bit, N43 need around 100CU to feed the data, that would need bigger die size than 250mm2 (another confirmed leaks). Plus GDDR7 won't be available until 2025...
I know 40CU sounds little with 192-bit GDDR6 memory and why do AMD release 32CU N44 with 128-bit GDDR6 memory? How about I show you potential AMD's mobile GPU roadmap next year:

2023 Model

N5/N6

Memory

2024 Model

N4P

Memory

7600M/XT

32/28 CU

8 GB

8600M/XT

32/28 CU

8 GB

8700M

36 CU

10 GB

8800M

40 CU

12 GB

7900M/XT

72/? CU

16GB

There are not many leaks on mobile RDNA4, but I think AMD is actually focusing on mobile platform with RDNA4. Remember notebook market is much bigger than desktop market like 80:20. If you remember 6700M and 6800M, that's the purpose of N43. And how many CU they are having?
You could compare die size between 8800M and RTX4080 cause they are both using 192-bit memory bus. N43's die is far smaller than AD104, thus AMD could potential undercut RTX4080, that's why NV is rumored to release RTX4070Ti to counter it.
AMD did mention that they are done with RDNA3 SKUs, so 7900M comes as surprise but I believe there are very few OEMs would adopt. I heard Alienware would be first OEM to use 7900M
If AMD is not announcing mobile N32 lineup with 7900M, then I am certain there won't have mobile N32 GPU at all cause AMD usually will stick two-year cadence with each generation.
Back to desktop GPU; there are 2 important price points AMD is missing: $299 and $399. NV already announced such products at these 2 price points. That's why I think AMD is reserving two price points with upcoming N43 and N44.

The points are way too long, the rest you could find in my old thread. Come back when AMD announces 7900M if you still don't believe..

adroc_thurston · Oct 12, 2023

Tigerick said:
These two dies are positioned like old Polaris

Ugh. No?

Tigerick said:
I hope nobody is buying GDDR7 BS

wot

jpiniero · Oct 12, 2023

Yeah I was under the impression that Micron or Samsung is starting HVM of GDDR7 in the first half next year. It might be too expensive for RDNA4''s price points, so it's not out of the question they would still use GDDR6.

adroc_thurston · Oct 12, 2023

jpiniero said:
It might be too expensive for RDNA4''s price points, so it's not out of the question they would still use GDDR6.

GDDR is inherently cheap.

TESKATLIPOKA · Oct 12, 2023

Tigerick said:
Ho, let's go point by point, shall we:

We know there are 2 monolithic dies made by N4P process.

These two dies are positioned like old Polaris, in case people are not familiar of it, have a look of old AT article here. Basically, they are die shrink of current RDNA3 with better power efficiency or lower price or both.

N43/N44/N48 are leak code-names but I don't see the point of N48, do you? N43 actually make sense cause the number is sitting below N32, maybe N48 is like damage control code-name.

You said N33 and N32 won't be discontinued, yet now you claim those 2 RDNA4 chips are basically a die shrink of RDNA3, so they should perform at least like them. In this case there is no reason to keep making N32 or N33.
The difference between N*2 and N*3 is the time development started and not performance level as far as I remember.

Tigerick said:
I hope nobody is buying GDDR7 BS cause with 192-bit, N43 need around 100CU to feed the data, that would need bigger die size than 250mm2 (another confirmed leaks). Plus GDDR7 won't be available until 2025...

Don't understand how you got 100CU, that is so wrong.
N32 has 256-bit bus + 19.5gbps for 60CU.
192-bit bus + 32gbps offers only 23% more BW, so It's good enough for only ~74CU.

Tigerick said:
I know 40CU sounds little with 192-bit GDDR6 memory and why do AMD release 32CU N44 with 128-bit GDDR6 memory? How about I show you potential AMD's mobile GPU roadmap next year:

2023 Model N5/N6 Memory 2024 Model

N4P

Memory
7600M/XT 32/28 CU 8 GB 8600M/XT 32/28 CU 8 GB
8700M 36 CU 10 GB
8800M 40 CU 12 GB
7900M/XT 72/? CU 16GB

There are not many leaks on mobile RDNA4, but I think AMD is actually focusing on mobile platform with RDNA4. Remember notebook market is much bigger than desktop market like 80:20. If you remember 6700M and 6800M, that's the purpose of N43. And how many CU they are having?

You could compare die size between 8800M and RTX4080 cause they are both using 192-bit memory bus. N43's die is far smaller than AD104, thus AMD could potential undercut RTX4080, that's why NV is rumored to release RTX4070Ti to counter it.

8600M has the same specs as 7600M, there is no reason to keep producing N33 If N44 has the same specs.
A 40CU GPU doesn't need 50% more BW compared to the one with only 32CU.
A 40CU N43 would never be on par with AD104, unless you could clock It >4GHz!
BTW, 40CU is a low amount for a 250mm2 N4P GPU.

Tigerick said:
AMD did mention that they are done with RDNA3 SKUs, so 7900M comes as surprise but I believe there are very few OEMs would adopt. I heard Alienware would be first OEM to use 7900M

If AMD is not announcing mobile N32 lineup with 7900M, then I am certain there won't have mobile N32 GPU at all cause AMD usually will stick two-year cadence with each generation.

Back to desktop GPU; there are 2 important price points AMD is missing: $299 and $399. NV already announced such products at these 2 price points. That's why I think AMD is reserving two price points with upcoming N43 and N44.

The points are way too long, the rest you could find in my old thread. Come back when AMD announces 7900M if you still don't believe..

AMD announced a cutdown 72CU N31 for mobile, which should be 2x faster than N33, leaving a big hole there. Instead of using N32, you think AMD will skip It and use N33 and N34 instead.
Now, you even claim $299 and $399 are the price points for desktop N43 and N44.
Then what about cutdown versions? $349 and $249?
Did you think what would happen to RX 7600 and RX 7700XT? Probably not.

You didn't provide any evidence, just your guesswork with some glaring mistakes, that's all.

SRP	2023	N5/N6	2024	N4P	2025/2026	N3E
$999	7900XTX 24GB 384-bit GDDR6	96 WGP			N51 24GB 384-bit GDDR7	210 WGP
$899	7900XT 20GB 320-bit GDDR6	84 WGP			N51L 20GB 320-bit GDDR7	?
$699					N52 16GB 256-bit GDDR7	140 WGP
$599					N52L ?	?
$499	7800XT 16GB 256-bit GDDR6	60 WGP			N53 12GB 128-bit GDDR7	70 WGP
$449	7700XT 12GB 192-bit GDDR6	54 WGP
$399			?	?
$299			?	?
$269	7600 8GB 128-bit GDDR6	32 WGP

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Lifer

Senior member

Senior member

Platinum Member

Platinum Member

Senior member

Lifer

Platinum Member

Platinum Member

Lifer

Platinum Member

Platinum Member

Diamond Member

Diamond Member

Lifer

Diamond Member

Diamond Member

Polaris System Specs​

Lifer

Platinum Member

Diamond Member

Senior member

Platinum Member

Lifer

Platinum Member

Platinum Member

Polaris System Specs