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

 

[adroc_thurston]

They've had a $1000 MSRP (ish) card since 2020. Didn't help!

Didn't win.
Also neither were halo. Halo is 650mm^2 area and up.

 

[gdansk]

Didn't win.
Also neither were halo. Halo is 650mm^2 area and up.

They're 'very high end'. And any actual gaming card over $1000 is - in my correct opinion - stupid.
Haloism works for Nvidia but it's a questionable goal for AMD.

 

[adroc_thurston]

Haloism works for Nvidia but it's a questionable goal for AMD.

It worked even better for ATi.
It's a tribal market. To lead the tribe you have to kill the previous chieftain.

 

[gdansk]

It worked even better for ATi.
It's a tribal market. To lead the tribe you have to kill the previous chieftain.

When they had leading software support with ... D3D 9?
I guess if AMD can pull off leading software and leading hardware at the same time it would work. But I considered the first impossible.

 

[adroc_thurston]

When they had leading software support with ... D3D 9?

They had the leading hardware.
D3D9 featureset compliance they were behind for a while.
Didn't matter an iota. A win is a win.

 

[SolidQ]

Didn't win.

Because DLSS/RT spam win

 

[insertcarehere]

Yeah but when everybody buys GeForce by default anyway, you will spend years and much effort to try to make Radeon good for those people and in the end you find that you sold 10 cards a quarter to them.

That's the sort of short term thinking that got AMD so far behind in the first place. Uni students that get exposed to GPU programming with consumer hardware (which is functionally CUDA stack at this point) become engineers who will have significant say, both explicit and implicit, on what hardware/software ecosystem they prefer to develop in. Nvidia 'making GeForce good for those people' is a big reason for their crushing lead in datacenter revenue right now.

 

[adroc_thurston]

Because DLSS/RT spam win

No, because they undersized the die.

That's the sort of short term thinking that got AMD so far behind in the first place. Uni students that get exposed to GPU programming with consumer hardware (which is functionally CUDA stack at this point) become engineers who will have significant say, both explicit and implicit, on what hardware/software ecosystem they prefer to develop in. Nvidia 'making GeForce good for those people' is a big reason for their crushing lead in datacenter revenue right now.

This narrative is a load of bull since CUDA as you know it is exclusively a Meta product.
All the actual GPGPU stuff went absolutely nowhere, right now GPUs in DC are just matrix accelerators.

 

[gdansk]

That's the sort of short term thinking that got AMD so far behind in the first place. Uni students that get exposed to GPU programming with consumer hardware (which is functionally CUDA stack at this point) become engineers who will have significant say, both explicit and implicit, on what hardware/software ecosystem they prefer to develop in. Nvidia 'making GeForce good for those people' is a big reason for their crushing lead in datacenter revenue right now.

AMD had early ghetto solutions on consumer hardware. E.g. early btc miners were tuned for AMD hardware, not Nvidia, and used OpenCL. But that wasn't a runaway success like ML.

People always blame AMD for this but shouldn't Khronos/OpenCL bear a lot of the blame? If every hardware vendor has a hard time maximizing performance... maybe something is wrong with it.

 

[SolidQ]

No, because they undersized the die.

Users don't care about die size, they care about perfomance.
Anyway i'm assuming NV with RTX5xxx series going heavy marketing Path Tracing(unless RDNA4 is surprising us with perf)/Neural Rendering, and again win with marketing

 

[adroc_thurston]

Users don't care about die size, they care about perfomance.

Die size always equates performance.
They didn't build big enough.

 

[SolidQ]

Die size always equates performance.

Yeah, i wonder why they didn't make a 80CU RDNA4 monolith size

 

[insertcarehere]

AMD had early ghetto solutions on consumer hardware. E.g. early btc miners were tuned for AMD hardware, not Nvidia, and used OpenCL. But that wasn't a runaway success like ML.

People always blame AMD for this but shouldn't Khronos/OpenCL bear a lot of the blame? If every hardware vendor has a hard time maximizing performance... maybe something is wrong with it.

Khronos group is a non-profit consortium that is composed of such disparate members as Apple, IKEA, Nvidia, Valve, Epic Games, Amazon, and ARM. Blaming them for AMD's current lack of traction is asinine when thats not their remit in the first place.

 

[gdansk]

Khronos group is a non-profit consortium that is composed of such disparate members as Apple, IKEA, Nvidia, Valve, Epic Games, Amazon, and ARM. Blaming them for AMD's current issues is asinine when that's never been anywhere near their remit in the first place.

Is it? That AMD would be better off ignoring the "open" standard is asinine. That's basically what all other arguments say: AMD should have made some alternative crap to compete with CUDA 15 years ago. It would have flopped even harder. AMD backed the wrong horse in 2008 but how was it supposed to know that trying to fix OpenCL 10 years ago would cause it to split and its creator to basically ignore it. But so it turned out to be.

 

[adroc_thurston]

Yeah, i wonder why they didn't make a 80CU RDNA4 monolith size

Because that's close to N48 and everything above was tiled anyway.

AMD should have made some alternative crap to compete with CUDA 15 years ago. It would have flopped even harder.

Well they did, HSA was a thing.
Just that GPGPU was a meme, APUs for compute were even more a meme and AMD had a long ass 5 year gap between proper GPGPU offerings.

 

[SolidQ]

Powercolor 9070XT

https://twitter.com/x/status/1875083081050448004

Fat card for middle

 

[reaperrr3]

Yeah, i wonder why they didn't make a 80CU RDNA4 monolith size

Making N48 with 80 CU would've been a bit more work than exactly doubling N44, and would also be somewhat bandwidth-starved unless they upgraded either the IF$ or used GDDR7, both of which would've been cost drivers.

If you meant why they didn't make a bigger chip above N48 with N31-like or higher specs once chiplets were cancelled:
Time to market, mostly.

The SKUs above N48 were all meant to be chiplet-based, and by the time they cancelled those, making another, bigger monolith above N48 would've taken too long.
Might've been a different story if they cancelled chiplet-RDNA4 earlier, then we might've gotten an "N31 specs + 24 CUs" RDNA4-based monolith, which could've spanked GB203, at least.
Though knowing Jensen, if he got info about that early enough, he would've probably just made NV aim higher with GB203 then, too.
There were rumors about 112 and 96 SM until the 84 SM info arrived, they probably down-specced GB203 during development when it became apparent they wouldn't need that many SM to beat AMD.

 

[marees]

Making N48 with 80 CU would've been a bit more work than exactly doubling N44, and would also be somewhat bandwidth-starved unless they upgraded either the IF$ or used GDDR7, both of which would've been cost drivers.

If you meant why they didn't make a bigger chip above N48 with N31-like or higher specs once chiplets were cancelled:
Time to market, mostly.

The SKUs above N48 were all meant to be chiplet-based, and by the time they cancelled those, making another, bigger monolith above N48 would've taken too long.
Might've been a different story if they cancelled chiplet-RDNA4 earlier, then we might've gotten an "N31 specs + 24 CUs" RDNA4-based monolith, which could've spanked GB203, at least.
Though knowing Jensen, if he got info about that early enough, he would've probably just made NV aim higher with GB203 then, too.
There were rumors about 112 and 96 SM until the 84 SM info arrived, they probably down-specced GB203 during development when it became apparent they wouldn't need that many SM to beat AMD.

No chip above N48 makes sense only if RDNA 5 is releasing soon (like in 1 year's time)

 

[gaav87]

Powercolor 9070XT

https://twitter.com/x/status/1875083081050448004

Fat card for middle

ye looks 2.5-3 slot

 

[inquiss]

No chip above N48 makes sense only if RDNA 5 is releasing soon (like in 1 year's time)

It makes sense not because rDNA 5 is coming soon (it's not) but because ROI on AI chips is bigger. Ditch the high end RDNA4 with uncertain ROI and move those people over to get a quicker cadence on AI chips with a (relatively) more certain ROI.

 

[yuri69]

No chip above N48 makes sense only if RDNA 5 is releasing soon (like in 1 year's time)

The summer '24 leaked roadmap shows RDNA5 to simply keep the trend of RDNA4 - leaving the segment previously covered by Navi 21 and Navi 31 out.

For AMD it makes zero sense to invest $$$ to try to compete with nV at higher price points. Consumers simply buy GeForce, like they used to do for decades. The buying decision is quite simple when one buys a ~$1000 consumer dGPU - the consumer wants the full feature set with the "standard experience". That's always been GeForce.

 

[SolidQ]

shows RDNA5 to simply keep the trend of RDNA4

If that leak real, UDNA still higher, than RDNA4, seems gonna compete with x8xx GF

 

[adroc_thurston]

No chip above N48 makes sense only if RDNA 5 is releasing soon (like in 1 year's time)

It makes sense when the halo tiled parts got canned.

 

[Kepler_L2]

No chip above N48 makes sense only if RDNA 5 is releasing soon (like in 1 year's time)

It's not.

 

[gaav87]

The summer '24 leaked roadmap shows RDNA5 to simply keep the trend of RDNA4 - leaving the segment previously covered by Navi 21 and Navi 31 out.

For AMD it makes zero sense to invest $$$ to try to compete with nV at higher price points. Consumers simply buy GeForce, like they used to do for decades. The buying decision is quite simple when one buys a ~$1000 consumer dGPU - the consumer wants the full feature set with the "standard experience". That's always been GeForce.

Leaked roadmap is from 1 year ago.
Rdna1 ->RDNA2 was 1 year 4 months gap not full 2 year cycle.