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

 

[SolidQ]

so 4 SKU's, seems gonna be same situation RX 5700XT vs R7, but faster RT

暫不推旗艦產品回應 NVIDIA，AMD RDNA 4 GPU 架構鎖定主流市場 - BenchLife.info

策略性的轉換，對現階段的 AMD 可能說是件好事。 AMD 在 RDNA 4 GPU 架構部分的規劃 […]

benchlife.info

AMD's plan for the RDNA 4 GPU architecture is actually very clear, namely Navi 48 and Navi 44. If it is further subdivided, Navi 48 will include R24D-E8, R24D-E6 and Navi 44's R25D. - Four products including P8 and R25D-P4 are in AMD's current planning.

At the same time, it is also confirmed that AMD will not launch flagship-level products like Radeon RX 7900 XTX in the RDNA 4 generation.

 

[marees]

so 4 SKU's, seems gonna be same situation RX 5700XT vs R7, but faster RT

暫不推旗艦產品回應 NVIDIA，AMD RDNA 4 GPU 架構鎖定主流市場 - BenchLife.info

策略性的轉換，對現階段的 AMD 可能說是件好事。 AMD 在 RDNA 4 GPU 架構部分的規劃 […]

benchlife.info

AMD's plan for the RDNA 4 GPU architecture is actually very clear, namely Navi 48 and Navi 44. If it is further subdivided, Navi 48 will include R24D-E8, R24D-E6 and Navi 44's R25D. - Four products including P8 and R25D-P4 are in AMD's current planning.

At the same time, it is also confirmed that AMD will not launch flagship-level products like Radeon RX 7900 XTX in the RDNA 4 generation.

I am very suspicious of benchlife

While kepler has been saying

N48xtx
N48xt
N48xl
N44xt

Benchlife says
N48xt
N48xl?
N44xt
N44xl

Also the 5th sku N44xl seems months behind the other 4 skus.

 

[marees]

I am very suspicious of benchlife

While kepler has been saying

N48xtx
N48xt
N48xl
N44xt

Benchlife says
N48xt
N48xl?
N44xt
N44xl

Also the 5th sku N44xl seems months behind the other 4 skus.

Ideally

($600) N48xtx = 64CU 8700xtx 16gb
($500) N48xt = 54CU 8700xt 16gb
($400) N48xl = 48CU 8600xt 12gb

($330) N44xt = 32CU 8500xt 16gb??
($280) N44xl = 28CU 8500 8gb

 

[Kepler_L2]

I am very suspicious of benchlife

While kepler has been saying

N48xtx
N48xt
N48xl
N44xt

Benchlife says
N48xt
N48xl?
N44xt
N44xl

Also the 5th sku N44xl seems months behind the other 4 skus.

192-bit version of N48 might come later if yields are really good.

 

[SolidQ]

($330) N44xt = 32CU 8500xt 16gb??

330 for 8500XT name?

 

[marees]

330 for 8500XT name?

They could name the whole thing a tier up. If they launch next year then RDNA 3 stocks would be gone & hence prices would be reasonable

But if launched this year, the way to go is name a tier up & price high

($600/$500) N48xtx = 64CU 8800xtx/8700xtx 16gb
($500/$400) N48xt = 54CU 8800xt/8700xt 16gb
($400/$350) N48xl = 48CU 8700xt/8600xt 12gb

($330/$300) N44xt = 32CU 8600xt/8500xt 16gb??
($280/$240) N44xl = 28CU 8500xt/8500 8gb

 

[Mahboi]

Frankly bad idea.
8700 XTX
8700 XT
8700
8600 XT (if clamshell doubling memory)
8600
would do great.

You don't want another "our x900 XT is the same performance as their xx70 Ti".
If anything, it'll make your cards look like tryhard weak hardware. Best to call it 7 at the top.

 

[Ghostsonplanets]

They could name the whole thing a tier up. If they launch next year then RDNA 3 stocks would be gone & hence prices would be reasonable

But if launched this year, the way to go is name a tier up & price high

($600/$500) N48xtx = 64CU 8800xtx/8700xtx 16gb
($500/$400) N48xt = 54CU 8800xt/8700xt 16gb
($400/$350) N48xl = 48CU 8700xt/8600xt 12gb

($330/$300) N44xt = 32CU 8600xt/8500xt 16gb??
($280/$240) N44xl = 28CU 8500xt/8500 8gb

Hopefully 8500 with 28 CU is sold at $199. No reason for a <=130mm² binned die with 6/8 GB VRAM to be sold for nearly $300.

 

[del42sa]

AMD announces unified UDNA GPU architecture — bringing RDNA and CDNA together to take on Nvidia's CUDA ecosystem

Two become one.

www.tomshardware.com

So, next gen GPUs will become unified. But what it means is that the UDNA GPUs will span from APUs to desktop to server/HPC products.

I guess it is a way of saying: Bye Bye RDNA, welcome to CDNA, without actually saying so

 

[Mahboi]

Hopefully 8500 with 28 CU is sold at $199. No reason for a <=130mm² binned die with 6/8 GB VRAM to be sold for nearly $300.

There hasn't been much reason in the GPU market in a long time.
I do hope that we see some, but I highly doubt that they won't try to make it at least $250.

 

[marees]

I guess it is a way of saying: Bye Bye RDNA, welcome to CDNA, without actually saying so

Could be the real reason the RDNA 4 is truncated. Just like RDNA 1 was a stop gap between GCN & RDNA 2 with RT

 

[Mahboi]

I also think it's likely that RDNA 4 was canned because they're going to reunite CDNA and RDNA and the workload was extensive enough that spending time on a "big" RDNA 4 didn't make much sense when there's already a lot to do for RDNA 5/URA-NUS (I will not call it UDNA until the name is final, I want to believe ).

 

[Ghostsonplanets]

There hasn't been much reason in the GPU market in a long time.
I do hope that we see some, but I highly doubt that they won't try to make it at least $250.

Then they'll be competing against cheaper priced RX 6600, 6650 and RX 7600.

 

[Mahboi]

Then they'll be competing against cheaper priced RX 6600, 6650 and RX 7600.

And they still need to get rid of that inventory.
It really is an awful market.

 

[Kepler_L2]

Could be the real reason the RDNA 4 is truncated. Just like RDNA 1 was a stop gap between GCN & RDNA 2 with RT

Not related at all.

 

[Hitman928]

Just my point of view on the matter, but if AI had not blown up like it did, I don’t think AMD would have switched back to a unified architecture. I think AI developers (AMD internal and customers) don’t like having to develop for 2 different architectures to support AMD’s product stack.

 

[adroc_thurston]

Bye Bye RDNA, welcome to CDNA, without actually saying so

The opposite, it's CDNA that sticks to a rather antique CU and ISA version.

Just my point of view on the matter, but if AI had not blown up like it did, I don’t think AMD would have switched back to a unified architecture. I think AI developers (AMD internal and customers) don’t like having to develop for 2 different architectures to support AMD’s product stack.

They're still gonna be separate just like NV DC and client microarchitecture slices are separate.

 

[gaav87]

If rdna4 is not launching end of OCT and its really CES 2025 its DOA period

Who is going to buy 8800xtx at 500$ even if its +-5% 7900xt/4080 when 5070 is going to be 599$? and most likely the same speed launched at the same time?
Only hope would be if 5070 649$ and still 12GB vram.
If the 5070 is 16GB and 599$ ~4080 perf.
Then GL&HF
I would buy 8800xtx if it launched next month/2months DAY ONE. Next year no thanks im going green first time since 2060s.

 

[dr1337]

Just my point of view on the matter, but if AI had not blown up like it did, I don’t think AMD would have switched back to a unified architecture. I think AI developers (AMD internal and customers) don’t like having to develop for 2 different architectures to support AMD’s product stack.

I think it might be a corporate type thing since they still sell RDNA cards as professional via w7800 and such. But I also feel like there can't possibly that many customers trying to run the same code on a radeon instinct card as well as the mi300/200, let alone users that would have issues with re-compiling.

It might be something organizational, perhaps the split development behind the scenes at AMD was too much overhead and counter-intuitive. If the RDNA team or CDNA team are working on the same problem but not sharing all resources then I can imagine how it might be difficult.

And doesn't ROCm make the architecture ambiguous from a code standpoint anyways? I don't get it from a software angle. AMD already has to develop and optimize two sets of drivers anyways even on the same architecture. For them to up and pivot like this surely comes from the inside.

 

[adroc_thurston]

And doesn't ROCm make the architecture ambiguous from a code standpoint anyways?

It does not, until SPIR-V support is mainlined, the entirety of ROCm is binary slices affixed for a particular gfx ISA.

If the RDNA team or CDNA team are working on the same problem but not sharing all resources then I can imagine how it might be difficult.

It's about abolishing the current ISA ghetto.
Compute parts will still have special sauce features with binary slices needed for them.

 

[dr1337]

the entirety of ROCm is binary slices affixed for a particular gfx ISA.

yeah but because ROCm exists the developers don't have to make those binary slices themselves. Even cuda works that way, no two architectures are fully alike and newer cards have features that aren't available on older ones.

You even seem to realize this in the last half of your reply lol.

 

[adroc_thurston]

yeah but because ROCm exists the developers don't have to make those binary slices themselves.

Which limits the support severely.

Even cuda works that way, no two architectures are fully alike and newer cards have features that aren't available on older ones.

CUDA has forward compatability and backwards goes pretty far too.
But yes, NV also adopted binary slices starting with H100.

 

[maddie]

If rdna4 is not launching end of OCT and its really CES 2025 its DOA period

This statement implies that the totality/majority of RDNA4 sales will occur in that time period. Do you realize this?

 

[ToTTenTranz]

AMD announces unified UDNA GPU architecture — bringing RDNA and CDNA together to take on Nvidia's CUDA ecosystem

Two become one.

www.tomshardware.com

So, next gen GPUs will become unified. But what it means is that the UDNA GPUs will span from APUs to desktop to server/HPC products.

Sounds like this is a good way to ensure the gaming architecture doesn't just die out from all the engineering allocation going to AI/compute.

 

[marees]

Sounds like this is a good way to ensure the gaming architecture doesn't just die out from all the engineering allocation going to AI/compute.

Exactly. This is it or bust for Radeon. MS about to release xbox o/s & stop hardware. PS & Valve are both on linux. If this doesn't work it could be the end of directX gaming for radeon