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

 

[gdansk]

Navi48 confirmed at 350mm2

Source?

 

[Saylick]

Source?

Allegedly 350mm2 with 53.9B xtors. It’s a dense boy. I wonder how they improved the density so much…

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

 

[CouncilorIrissa]

Allegedly 350mm2 with 53.9B xtors.

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

These are seemingly eyeball measurments still. Not "confirmed" by any means.
xtor counts are, however.

 

[gdansk]

Allegedly 350mm2 with 53.9B xtors. It’s a dense boy. I wonder how they improved the density so much…

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

Citing a Chinese forum post that at least references an Anandtech forum post...
We're getting really circular here

 

[Kepler_L2]

Allegedly 350mm2 with 53.9B xtors. It’s a dense boy. I wonder how they improved the density so much…

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

David Wang is an absolute madman, targeting 154 MTr/mm² and 3.46 GHz at the same time.

 

[Saylick]

David Wang is an absolute madman, targeting 154 MTr/mm² and 3.46 GHz at the same time.

Literally David vs Goliath (Nvidia).

 

[Tup3x]

That's encouraging that AMD is asking for feedback on pricing. Hopefully they listen to Steve from HWUB - aggressive undercutting with high supply would win them a LOT of mindshare right now. Especially because anything over RTX 4060 level is basically unobtanium or stupidly priced.

I think it's a bad sign that company doesn't know how to price their product. They know how their product performs relative to competition. They can guess how RTX 5070 would perform. If this means that they aren't going to set the price too high this time, then this is a good thing. That being said, they should have figured things out by themselves...

 

[Ghostsonplanets]

It's probably x8 PCIe since N48 is x16. And yeah it seems to lack media encoder.

If it's x8, then the 9060 will probably be based in N44. Lack of media encoder is relatively bad, but given AM5 CPUs now have iGPUs with Media Engine, it's not as bad of a situation as with Navi 24

RDNA4 has nearly 50% improved performance per CU.

Seriously? That's extremely impressive uplift gen on gen.

David Wang is an absolute madman, targeting 154 MTr/mm² and 3.46 GHz at the same time.

3.46GHz? And with such density? Did AMD came into contact with some superior being technology😳?

Joke aside, I believe the fundamentals of RDNA 3 were always there to be explored and refined further. So it wouldn't surprise me. The cross-pollination of Zen and RDNA teams R&D also are proving to be extremely effective.

 

[gaav87]

Citing a Chinese forum post that at least references an Anandtech forum post...
We're getting really circular here

Ye i just noticed that

 

[adroc_thurston]

That's extremely impressive uplift gen on gen.

No that's normal (sort of).

3.46GHz? And with such density? Did AMD came into contact with some superior being technology😳?

That's for N3e parts.

Joke aside, I believe the fundamentals of RDNA 3 were always there to be explored and refined further. So it wouldn't surprise me.

This is not related to RDNA3 in any way, shape or form.

 

[Kepler_L2]

3.46GHz? And with such density? Did AMD came into contact with some superior being technology😳?

I don't know if they actually managed to hit, so far leaked boost clocks are lower but advertised numbers aren't always accurate.

 

[maddie]

I think it's a bad sign that company doesn't know how to price their product. They know how their product performs relative to competition. They can guess how RTX 5070 would perform. If this means that they aren't going to set the price too high this time, then this is a good thing. That being said, they should have figured things out by themselves...

There must be some furious internal battles over pricing happening now.

 

[Kepler_L2]

That's for N3e parts.

Nope that was the target for N48

 

[GTracing]

If the die is 350mm^2 and the 9070 XT performs like a 5070 ti, RDNA4 would actually have better PPA than Blackwell.

AMD has come a long way from when the 232mm^2, 150W, 256bit, RX 480 was losing to the 200mm^2, 120W, 196bit GTX 1060.

 

[adroc_thurston]

Nope that was the target for N48

I know but in this case it's more of a stretch goal.

If the die is 350mm^2 and the 9070 XT performs like a 5070 ti, RDNA4 would actually have better PPA than Blackwell

Yeah 350 or a bit less and it's a bit faster than that.
Mind you, it's also a lot less bandwidth.

 

[GodisanAtheist]

AMD asking for reviewer opinions and feedback on pricing is oddly refreshing for a corpo.

They know what they don't know (their ground level perception) and they're doing the corpo thing of collecting data and feedback.

My guess is they're going to take the ground floor pricing reviewers suggest and +$50 that and run with it to make the bean counters internally happy while also keeping reviewers from absolutely skewering the review units on pricing.

 

[exquisitechar]

If the die is 350mm^2 and the 9070 XT performs like a 5070 ti, RDNA4 would actually have better PPA than Blackwell.

AMD has come a long way from when the 232mm^2, 150W, 256bit, RX 480 was losing to the 200mm^2, 120W, 196bit GTX 1060.

And it’s using GDDR6 instead of GDDR7. The caches got really beefed up for sure.

 

[adroc_thurston]

And it’s using GDDR6 instead of GDDR7. The caches got really beefed up for sure.

Not really beefed up, just different in some ways.

 

[GTracing]

And it’s using GDDR6 instead of GDDR7. The caches got really beefed up for sure.

Right, RDNA4 only has better PPA if you take that into account (and that's assuming these rumors are true). Otherwise Blackwell is slightly better.

 

[SolidQ]

I don't know if they actually managed to hit, so far leaked boost clocks are lower but advertised numbers aren't always accurate.

Maybe they left it for user OC?

 

[coercitiv]

There must be some furious internal battles over pricing happening now.

They've been working hard for months now, but have only agreed on the last two digits.

I think it's a bad sign that company doesn't know how to price their product.

If I were to play devil's advocate, I could argue they're asking for feedback from reviewers because they're doing a sort of sanity check. It's like getting free analytics, not the greatest quality assessment but still quite useful.

I'll refrain from writing down my actual opinion, because I've been writing on the end of this post for quite a while and it always ends in words that break forum rules.

 

[basix]

Allegedly 350mm2 with 53.9B xtors. It’s a dense boy. I wonder how they improved the density so much…

The did not gain too much density at all. N31 GCD sits at ~149 Mtr/mm2 and that is N5 and not including dense Infinity Cache (but also not so dense GDDR6 PHY). As RDNA4 is using a slightly denser process node (N4), it is realistic to achieve ~150 Mtr/mm2.

AMD asking for reviewer opinions and feedback on pricing is oddly refreshing for a corpo.

They know what they don't know (their ground level perception) and they're doing the corpo thing of collecting data and feedback.

You could also argue, that AMD is asking for the maximum price tag they get away with, where they still get positive reviews

 

[adroc_thurston]

As RDNA4 is using a slightly denser process node (N4), it is realistic to achieve ~150 Mtr/mm2.

Uh, nope.
GPUs have a ton of shoreline stuff and a ton of SRAM which doesn't really shrink much.

 

[sandorski]

Did AMD delay just to build Stock? That might be their main goal, to enter the Market with a reasonably priced product with high Supply where competitor low Supply limits competition. Temporarily they can grab Marketshare and even take some Sales that were going to go NVidia's way.

 

[basix]

Uh, nope.
GPUs have a ton of shoreline stuff and a ton of SRAM which doesn't really shrink much.

Also N31 has a lot of shoreline stuff, although GDDR6 PHY traded with Infinity Fanout Links.
And N31 does not feature a huge LLC on the GCD and SRAM achieves highest transistor density, higher than any logic IP. So if you transfer that huge SRAM array to the main Die. your density should increase. Sure, the bigger GDDR6 PHY will offset that again.

I am pretty convinced that the transistor density of the SEs etc. will be very similar to RDNA3. So no black magic transistor density on RDNA4. That is my point.