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

 

[misuspita]

Yeah, it may be rubbing their noses in their approach. "this is the way you deal with us now". I'd frankly be doing something like this, but then, I'm not the CEO of a multi billion dollar company. 😁

 

[GodisanAtheist]

AMD has never been able to supply the Laptop market/OEMs properly, it's just never been an issue since no OEMs ever actually wanted AMD other than their one value proposition "see Intel is not a monopoly" laptop edition.

AMD is competing with Apple and Nvidia, two of the largest companies on earth, for fab space at TSMC, during an AI gold rush.

AMD has that business poverty mindset, it's DC is making it money right now, so today it is a DC company. When the DC market cools down, it will be a *whatever is out most profitable product today* company.

Simple as.

 

[Mahboi]

Exactly this. ^
I wish people stopped trying to find excuses for AMD.
They were always behind on software, on docs, on support, on volume. They have been for at least 15 years, if not 20, if not forever.
They were the N°2 in everything since forever and didn't spend as much on extra SW or support people as NV/Intel.

They're just bad and need to grow better on these fields, the end.
Kraken being on SS4 is probably a good sign of understanding the problem and at least fixing volume.
Streamlining their docs and having a centralised source for information/dev comms would be a logical next step. I still have a feeling that most devs still don't really understand how to implement ROCm/HIP/FSR and could use a forum to talk to each other and AMD devs.

 

[Thunder 57]

Exactly this. ^
I wish people stopped trying to find excuses for AMD.
They were always behind on software, on docs, on support, on volume. They have been for at least 15 years, if not 20, if not forever.
They were the N°2 in everything since forever and didn't spend as much on extra SW or support people as NV/Intel.

They're just bad and need to grow better on these fields, the end.
Kraken being on SS4 is probably a good sign of understanding the problem and at least fixing volume.
Streamlining their docs and having a centralised source for information/dev comms would be a logical next step. I still have a feeling that most devs still don't really understand how to implement ROCm/HIP/FSR and could use a forum to talk to each other and AMD devs.

You could say that back when they were dependant on Via/Nvidia chipsets, but not longer.

 

[Ghostsonplanets]

Kraken being on SS4 is probably a good sign of understanding the problem and at least fixing volume.

Kraken is TSMC N4P (And maybe 4C in the future).

The SF4 product was the $299 - 399 product segment Sonoma Valley (4x Z5C + 1 RDNA 3.5 WGP) and I'm not sure it even exists anymore as AMD has cancelled a lot of products and the Samsung Foundry x AMD rumblings have quiet down lately.

Regardless, Kraken will be a volume part and should enjoy significant adoption from OEMs as it will be the only $799 x86 Copilot+PC part in the market next year. Intel doesn't has a 40 TOPs solution in this price range and QCOM is QCOM.

Everything depends on OEMs. But AMD expects an increase in design wins starting next year per Lisa Su calls.

 

[Mahboi]

Kraken is TSMC N4P (And maybe 4C in the future).

The SF4 product was the $299 - 399 product segment Sonoma Valley (4x Z5C + 1 RDNA 3.5 WGP) and I'm not sure it even exists anymore as AMD has cancelled a lot of products and the Samsung Foundry x AMD rumblings have quiet down lately.

Oh boy.

 

[adroc_thurston]

and I'm not sure it even exists anymore

It does.

AMD has cancelled a lot of products

They didn't?

 

[marees]

So AMD seems to have added more focus to software now::

1. UDNA vision
2. Universal ROCM to now work on all Radeon gaming cards
3. FSR 4 with AI/ML upscaling
4. More testing of drivers for games ??

What are the implications for RDNA 4 release

1. Delay to jan next year to get software ready ?
2. Reveal just before 5080/5090 to spoil Nv's party ?
3. Sneak attack before winter holiday season (using boards of 7700xt ??) ?
4. Wait & watch for battlemage gpu/lunar lake igpu & act accordingly ?

 

[SteinFG]

Delay to jan next year to get software ready ?

Feels like ROCm support will still lag after release, so delaying GPU launch is not gonna help

Reveal just before 5080/5090 to spoil Nv's party ?

Those cards don't even reach 5080-level performance (Unless there's literally no PPC and Clock improvements between lovelace and blackwell), so they're in different tiers. At most, AMD's card will compete with 4070 Ti Super, less likely 4080 super, by most recent rumors

Sneak attack before winter holiday season (using boards of 7700xt ??) ?

What even??? You think AMD has a secret factory that stockpiled Navi 32 PCBs? Absurd. It'll be a usual lauch.

 

[del42sa]

So AMD seems to have added more focus to software now::

1. UDNA vision
2. Universal ROCM to now work on all Radeon gaming cards
3. FSR 4 with AI/ML upscaling
4. More testing of drivers for games ??

What are the implications for RDNA 4 release

1. Delay to jan next year to get software ready ?
2. Reveal just before 5080/5090 to spoil Nv's party ?
3. Sneak attack before winter holiday season (using boards of 7700xt ??) ?
4. Wait & watch for battlemage gpu/lunar lake igpu & act accordingly ?

RDNA was a disaster, so why not try again and better. It wouldn't be a problem if AMD hadn't repeated it for the several times and with several failed atempts (Fiji with HBM/Vega/ RDNA3) I don't believe they'll end up with any significant market share. Still the same misery or excuses why it's not possible, etc.

AMD/RTG already had their chances and totally REDACTED up... plus the PC gaming area is declining more and more in favor of console play, so if that market ends, it won't be a tragedy for the company. Gaming graphics have high costs yet low margins.... (unless you´re Nvidia) So it makes sense to focus on the markets that bring in the most money and that´s not gaming GPUs.

Let's be realistic, plus it's AMD , so disappointment is guarantee

 

[Mahboi]

RDNA was a disaster,

lolwut

same misery or excuses why it's not possible, etc.

Are you feeling alright?

AMD/RTG already had their chances and totally fucked up... plus the PC gaming area is declining more and more in favor of console play,

Consoles are literally dying. Exclusives stream out of them every day, everything will end up on PC.

so if that market ends, it won't be a tragedy for the company. Gaming graphics have high costs yet low margins....

Consoles have tiny margins and gaming GPUs + compute market has a zillion times better margin potential than even server CPUs.

(unless you´re Nvidia) So it makes sense to focus on the markets that bring in the most money and that´s not gaming GPUs.

They're literally unifying gaming GPUs and compute GPUs just like Nvidia. What the hell are you on about?

Let's be realistic, plus it's AMD , so disappointment is guarantee

This is either low quality bait or you wrote this while high. Everything is wrong.

 

[Mahboi]

Feels like ROCm support will still lag after release, so delaying GPU launch is not gonna help

Those cards don't even reach 5080-level performance (Unless there's literally no PPC and Clock improvements between lovelace and blackwell), so they're in different tiers. At most, AMD's card will compete with 4070 Ti Super, less likely 4080 super, by most recent rumors

That's not how that works. Public attention span is a thing. If AMD can get some media noise going out of NV's yard and into AMD's, it works, competitive or not.
Reminder that when FSR 3 was launched and worked on everything including Ampere, Nvidia was unusually quick at releasing their unfinished AI denoiser thing that was clearly not ready. Because they wanted to drive the media circus away from asking intelligent questions like "why can DLSS 3 supposedly not work on 3000 series GPUs according to Nvidia but somehow FSR 3 can do it just as well as on the 4000s?".

The media drives a lot of public image.

 

[del42sa]

lolwut

Care to explain, how RDNA actually helped AMD gain any significant dGPU market share ?

Are you feeling alright?

Me ? yes, thanks for asking

Consoles are literally dying. Exclusives stream out of them every day, everything will end up on PC.

yes , that´s why AMD is working on new chips for both console. Lol

They're literally unifying gaming GPUs and compute GPUs just like Nvidia. What the hell are you on about?

Forgot about GCN ? They have been in that shoes already and result is ?

This is either low quality bait or you wrote this while high. Everything is wrong.

Sure, whatever

 

[Thunder 57]

Care to explain, how RDNA actually helped AMD gain any significant dGPU market share ?

Because stupid misinformed people bought 3050's instead of much more performant RX 6600's at basically the same price.

 

[Mahboi]

yes , that´s why AMD is working on new chips for both console. Lol

I don't think you understand how it works. SONY asks AMD to provide chips. AMD sells it to them.
AMD didn't go to Mark Cerny and say "we wanna invest in the console business". They just got ordered a bunch of chips and make money off it. SONY's investment in the console business engages AMD in no way shape or form.
Again are you feeling ok? You're failing at the most basic logic here.

Forgot about GCN ? They have been in that shoes already and result is ?

GCN is literally the basis for CDNA 1/2/3. CDNA 3/MI300 just happens to be the biggest, baddest chip in the entire world. Nothing even comes close to it, NV doesn't even try to equal it, they just focus on SW support and NVLink to keep it at arms' length.

So the result is pretty damn great for what it is.
Again you're failing at the most basic logic, several times in the same sentence actually, since you simultaneously throw random quips at AMD for making RDNA and at them for having made GCN.

It's like i'm back on Plebbit and hearing an Nvidia bot finding the most contrived excuses to claim that AMD can do no right no matter how right things go for them.

Sure, whatever

I'm not even sure what the point of that post was tbh, having read it 3 times now. Just feels like I'm reading Reddit.

In any case it's going to be a pretty damn boring 2 years with RDNA 4, hopefully MI350 gets something fun going.

 

[del42sa]

Are you seriously coming on the anandtech forums and going: "LOL RDNA IS SUCH A FAILURE LAWL LAWL AMD WAS IN SUCH A BETTER PLACE BACK IN THE VEGA/POLARIS YEARS" ?
This isn't Reddit.

you did not answer my question

GCN is literally the basis for CDNA 1/2/3. CDNA 3/MI300 just happens to be the biggest, baddest chip in the entire world. Nothing even comes close to it, NV doesn't even try to equal it, they just focus on SW support and NVLink to keep it at arms' length

thank you for enlightening elaborate about AMD success

 

[SamMaster]

so, it's people to blame, that AMD GPUs are failing ...

Partially. There are / have been instances where AMD / ATi had the superior product, even if just a few segments, and didn't fare well in sales.

 

[Mahboi]

you did not answer my question

You're the one making insane claims and you're the one with the burden of proof here.
I don't have to answer anything, you do.

thank you for enlightening elaborate about AMD success

....what?

Ok you are definitely drunk or smth.

 

[DAPUNISHER]

@del42sa

Stop with the AMD bashing and flame bait in an AMD thread. You got a ban bingo space filled in for using profanity too. Keep it civil, keep it on topic, please.

Mod DAPUNISHER

 

[SolidQ]

bought 3050's instead of much more performant RX 6600's at basically the same price.

not only this.
4060 vs 6700XT
4060ti 8 vs 7700XT
4060ti 16 vs 7800XT
but people will continue buy trash from NV even if 8600XT with 7700XT perfomance around 330$

What is point take 4060?

 

[SteinFG]

What is point take 4060?

4060 is usually 30 dollars cheaper though, that's all

 

[ToTTenTranz]

So is FSR4 only for UDNA and SoCs with NPUs, or is there a chance we'll see FSR4 on RDNA4 dGPUs?

 

[Mahboi]

lolwut?
Obviously if it's coming out near PS5 Pro, it's at least RDNA 4.
I'm tempted to believe that it'll also work with RDNA 3 albeit with worse perf. RDNA 3 already had Bfloat 16 and INT8.

 

[Kepler_L2]

So is FSR4 only for UDNA and SoCs with NPUs, or is there a chance we'll see FSR4 on RDNA4 dGPUs?

Probably for anything that supports WMMA.

 

[Mahboi]

Good, I was afraid of taking out the 7900 XT out of my mobo, last I had to change GPUs, I pulled the PCI-E x16 knob thing that locks cards so hard I thought I would break the mobo...
It's staying in there until it dies now.