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

 

[Timorous]

I told you guys that cube wasn't solid.

It isn't now that is for sure!

 

[gdansk]

It isn't now that is for sure!

The Sapphire cube has disintegrated. Existence is futile.

 

[Timorous]

The Sapphire cube has distintegrated. Existence is futile.

They still have 9070s in stock at 539.99 so the XT cube may be defeated but the plain one is still hanging in there.

 

[In2Photos]

Just hopping on for the first time today. Does anyone know if AMD was able to sell all 3 units today or is it going to take some time?

 

[Kronos1996]

Just hopping on for the first time today. Does anyone know if AMD was able to sell all 3 units today or is it going to take some time?

I have it on good authority that 2 of them sold but the 3rd person put it back on the shelf since AMD doesn’t have a halo card.

 

[Mahboi]

More news is coming

https://twitter.com/x/status/1897741520200962308

PSSR confirmed to be a partly shared tech. Wonder how that phone/Samsung partnership for FSR is going to turn out too.
It's strange, after years of dark limbo with RDNA 3, it's like a long-awaited birthday party finally comes our way.

 

[Mahboi]

Yeah they might have juuuust pushed it a lil' smidge too hard for nothing.

 

[Ranulf]

Yeah they might have juuuust pushed it a lil' smidge too hard for nothing.

Pretty much standard for AMD since Polaris or even the r9 290 cards. Good news though. Power savings at no reall perf loss.

 

[Mahboi]

Pretty much standard for AMD since Polaris or even the r9 290 cards. Good news though. Power savings at no reall perf loss.

Honestly feel like they ate their own foot here. Had the XT been 260W and the non 220W, the reviews would have been far more positive yet. Oh well.

 

[GTracing]

Pretty much standard for AMD since Polaris or even the r9 290 cards. Good news though. Power savings at no reall perf loss.

Yeah, both vendors add in a safety margin. It's pretty typical for GPUs to have a lot of undervolting headroom. Even the 5080 can be undervolted to 260W with roughly stock performance

 

[GodisanAtheist]

The Sapphire cube has distintegrated. Existence is futile.

- The cube encountered Species 8472 total cards produced.

 

[Mahboi]

Afraid of the answer I might get but what is this Cube story?

 

[Cstops]

Afraid of the answer I might get but what is this Cube story?

It’s in relation to this on an earlier post: A cube of 9070s in stock at a particular outlet.

https://videocardz.com/newz/uk-retailer-confirms-stock-of-1000-units-of-a-single-radeon-rx-9070-xt-model-at-msrp

 

[GTracing]

Afraid of the answer I might get but what is this Cube story?

It's referencing the borg cube from star trek.

 

[Mahboi]

It’s in relation to this on an earlier post: A cube of 9070s in stock at a particular outlet.

https://videocardz.com/newz/uk-retailer-confirms-stock-of-1000-units-of-a-single-radeon-rx-9070-xt-model-at-msrp

They stole all of Europe's supply.
We must rearm the Ships of the Line and end their evil.

 

[HurleyBird]

Let's hope AMD is gonna announce GDRR7 version with +15% perf

It feels like there's tension here because I can see us getting a G7 version or a 32GB version in the future, but I have a hard time seeing both. It would create a confused product stack, while the main thing G7 would do for a theoretical 32GB "ghetto prosumer" SKU is balloon cost and/or shrink margins, which doesn't make much sense when whole the whole selling point of that SKU would be capacity/$.

 

[blckgrffn]

It feels like there's tension here because I can see us getting a G7 version or a 32GB version in the future, but I have a hard time seeing both. It would create a confused product stack, while the main thing G7 would do for a theoretical 32GB "ghetto prosumer" SKU is balloon cost and/or shrink margins, which doesn't make much sense when whole the whole selling point of that SKU would be capacity/$.

More likely a new top sku with 24GB of GDRR7 to match the Blackwell Super refresh likely to do the same. G7 has memory chips in multiples of threes coming in the pipeline so todays 16GB cards can be next years 24GB cards.

 

[fastandfurious6]

View attachment 119129

Yeah they might have juuuust pushed it a lil' smidge too hard for nothing.

yeah defaults are set on safe side, can't have online posts complaining random crashes lol

this is big, 9700XT @ 260W with perf close to desktop RTX5080 @ 350-400W

top-end laptop gpu W is around 200W, total system 'thermal/W threshold' around 250-300W

so RDNA5 mobile will likely have similar perf up to desktop RTX5070 - 5080

i.e. Medusa Halo on a heavier SKU as well

 

[Kronos1996]

My technically astute friend, who annotates die shots, thinks its unlikely Navi 48 has the PHY’s to support GDDR6 and GDDR7. If it does there should be physical evidence, which remains to be seen without better pictures. They’re very different interfaces so it could be a lot of wasted silicon to add both. Unless you had a very good reason.

I’m skeptical that was ever planned for the small monolithic Navi 44 and therefore Navi 48. Probably was for the big chiplet designs. We should get more info when high resolution die shots are available for study. Only reason I could see them doing this is for Pro cards. AMD is notoriously stingy with silicon.

Edit: Added this die annotation courtesy of High Yield. Looks like pretty standard GDDR6 PHY’s. Not sure there any evidence of GDDR7 support but I’m by no means an expert.

 

[blckgrffn]

My technically astute friend, who annotates die shots, thinks its unlikely Navi 48 has the PHY’s to support GDDR6 and GDDR7. If it does there should be physical evidence, which remains to be seen without better pictures. They’re very different interfaces so it could be a lot of wasted silicon to add both. Unless you had a very good reason.

I’m skeptical that was ever planned for the small monolithic Navi 44 and therefore Navi 48. Probably was for the big chiplet designs. We should get more info when high resolution die shots are available for study. Only reason I could see them doing this is for Pro cards. AMD is notoriously stingy with silicon.

That would also make perfect sense to me… is it really impossible they wouldn’t update/change the memory controllers for a midcycle refresh?

Improbable I’ll grant for sure. But given where the market is and the likely possibility of a validated G7 memory controller for Pro cards existing, it really couldn’t happen?

 

[moonbogg]

It’s in relation to this on an earlier post: A cube of 9070s in stock at a particular outlet.

https://videocardz.com/newz/uk-retailer-confirms-stock-of-1000-units-of-a-single-radeon-rx-9070-xt-model-at-msrp

No way an employee took the time to make a solid cube of boxes just for a pic. That cube wasn't a cube, but a wall. Just ain't no way. Besides, we saw there wasn't enough stock for people to be making solid cubes anyway. This was a wall launch, not a cube launch.

 

[gdansk]

No way an employee took the time to make a solid cube of boxes just for a pic. That cube wasn't a cube, but a wall. Just ain't no way. Besides, we saw there wasn't enough stock for people to be making solid cubes anyway. This was a wall launch, not a cube launch.

The overachieving marketing employee constructed a hypercube of graphics cards. Imagine all the n-dimensional Navi 48s.

 

[Hans Gruber]

It will be interesting to see how much room the GDDR6 memory chips have for overclocks.

 

[marees]

No way an employee took the time to make a solid cube of boxes just for a pic. That cube wasn't a cube, but a wall. Just ain't no way. Besides, we saw there wasn't enough stock for people to be making solid cubes anyway. This was a wall launch, not a cube launch.

It was a solid cube. The store spent hours arranging a cube.

All stores were asked by AMD to post stock pics & upload to social media

Demand went thru the roof because nv was MIA

 

[gdansk]

It will be interesting to see how much room the GDDR6 memory chips have for overclocks.

Samsung claim to be sampling 24gbps GDDR6 chips. But I'll believe it when I see it.
For other GDDR6 chips it isn't unusual to get 22gbps stable enough. That would be a 10% memory overclock.