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

 

[soresu]

Can you read Bluesky posts without an account?

Yes, and in proper order too.

You can read some posts on Twitter without an account, but the order of posts is completely garbled for some reason when you click on the profile name.

 

[igor_kavinski]

they really oughta strap GDDR7 onto it and launch as 9080XT.

Not at the prices Nvidia must be paying the memory makers.

 

[adroc_thurston]

Not at the prices Nvidia must be paying the memory makers.

AMD literally sleeps with Hynix. They're basically married and have been so for almost 2 decades.

 

[GodisanAtheist]

AMD literally sleeps with Hynix. They're basically married and have been so for almost 2 decades.

- Is the magic gone and the spark dead? Maybe Hynix accepted NV's indecent proposal?

 

[adroc_thurston]

- Is the magic gone and the spark dead? Maybe Hynix accepted NV's indecent proposal?

Nope, MI325X is the thing with the meanest Hynix stacks on the market.

 

[CastleBravo]

Maybe something in Jensen's presentation worried the heck out of them and they are trying to offer some cool driver feature. Could it be the Transformer Model that spooked AMD? That would be my best guess.

Conspiracy theory: They bought some time on one of the supercomputers they helped build with the MI300 and frantically training their model to get a leg up on Jensen.

I doubt they are going to gut and rebuild FSR4 on a new model before they release RDNA4, but the DLSS4 transformer model might be what has them spooked. They probably want to polish FSR4 so that it is as close as possible to DLSS3. Even at that level of polish (unlikely I'm afraid), DLSS4 on 1/2 resolution scaling would still be able to match FSR4 image quality at 2/3 scale erasing some or maybe all of the performance to price advantage they might have. If FSR4 ends up looking more like DLSS2, then AMD upscaling will be just as far behind Nvidia as they were when Ampere was released.

 

[igor_kavinski]

Can AMD release something like the current RDNA4 die overclocked to the gills with zero overclocking headroom left, with 32GB VRAM? Is it possible to have a paper launch of such a product in two months?

 

[adroc_thurston]

Can AMD release something like the current RDNA4 die overclocked to the gills with zero overclocking headroom left, with 32GB VRAM? Is it possible to have a paper launch of such a product in two months?

They don't have to.
Just ship a GDDR7 board for it.
But that's later if they ever do that.

Is it possible to have a paper launch of such a product in two months?

If you're hoping they're scrambling to drop a better product because of magical Nvidia rays making them panick, I have bad, truly horrific news for you.

 

[soresu]

AMD literally sleeps with Hynix. They're basically married and have been so for almost 2 decades.

Did it start because of HBM, or before that?

 

[adroc_thurston]

Did it start because of HBM, or before that?

Way way way before that, they brough GDDR5 to market jointly too.

 

[poke01]

they really oughta strap GDDR7 onto it and launch as 9080XT.

this would be great. add it to the roadmap AMD

 

[adroc_thurston]

this would be great.

Well it would be just a hilarious accident.
I don't think AMD ever expected to ever get close to x80 comp this product cycle.

 

[Gideon]

Well it would be just a hilarious accident.
I don't think AMD ever expected to ever get close to x80 comp this product cycle.

You yourself said that Navi 44 and 48 have no GDDR7 support in memory controller,.

Why are you hyping up a (seemingly) impossible thing?

 

[itsmydamnation]

You yourself said that Navi 44 and 48 have no GDDR7 support in memory controller,.

Why are you hyping up a (seemingly) impossible thing?

could have sworn he said the exact opposite ?

 

[adroc_thurston]

You yourself said that Navi 44 and 48 have no GDDR7 support in memory controller,.

I've said the exact opposite.

Why are you hyping up a (seemingly) impossible thing?

Try getting your eyes checked.

 

[gaav87]

Anyway FMF 2.1 is, a thing slightly better than 2.0.

 

[lucasworais]

they really oughta strap GDDR7 onto it and launch as 9080XT.

With same amount of cores and clocks as their x70 card? 9070 XTX would suit better or 9770 XT even

 

[SolidQ]

9070 XTX

9070XT AI MAX!

from where that N4C rumor is came?

https://twitter.com/x/status/1883035137799266641

 

[Timorous]

9070XT AI MAX!

from where that N4C rumor is came?

https://twitter.com/x/status/1883035137799266641

This is AMD. You think they will keep to a naming convention. If they did revive 4c it would clearly be called something like Radeon Multi GPU 9 AI Max XTX ultra

 

[reaperrr3]

Not at the prices Nvidia must be paying the memory makers.

Yeah, AMD would have to pay more per chip due to lower volume.

But if full N48 at ~330W PT with 28Gbps G7 is faster than the 5070 Ti and close to 5080, I don't see why it shouldn't be profitable at something like $799.

$799 - 9080 XT -> 64 CUs, 2.65 base, 3.1 max, 330W, 16GB G7 28Gbps -> positioned as cheaper 5080 near-equivalent (or "faster than 5070Ti for nearly same price")
$649 - 9070 XT -> 64 CUs, 2.4 base, 2.97 max, 265W, 16GB G6 20Gbps -> positioned as cheaper and more efficient 5070 Ti equivalent
$549 - 9070 -> 56 CUs, 2.07 base, 2.52 max, 220W, 16GB G6 18Gbps -> positioned as faster, more efficient + bigger VRAM alternative to 5070

A line-up like that could probably work.

The bigger question is whether NV simply bought ALL current GDDR7 volume or not and if AMD could get any meaningful volume on short notice at all.

 

[MrTeal]

$150 for an extra 130MHz and some membw?

Dibs out.

 

[adroc_thurston]

$150 for an extra 130MHz and some membw?

Dibs out.

Hey that's like 10% perf (generously). Good nuff.

 

[Win2012R2]

The bigger question is whether NV simply bought ALL current GDDR7 volume or not and if AMD could get any meaningful volume on short notice at all.

AMD placing orders now will be getting it next year... (sauce: me)

 

[IEC]

If 5090 idle power draw is any indication, GDDR7 doesn't exactly sip power.

Making a midrange die with commodity GDDR6 is the right move. Launching after nVidia is also the right move historically. Pricing remains to be seen...

... though knowing AMD, they'll screw up on that one.

 

[soresu]

If 5090 idle power draw is any indication, GDDR7 doesn't exactly sip power.

The coming capacitorless DRAM designs should be a godsend to idle power draw given they only need to be refreshed every 400+ seconds, not to mention finally enabling multilayer devices like 3D/V NAND flash memory has used for the last decade.