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]

to cover 500 million in development costs?

oh it doesn't cost even remotely as much.

 

[beginner99]

Anyway, 9070 and 9070XT are 1440p cards

So a 7900 XTX and 4080 are 1440p cards as well?

Sel 9700 XT for 450$ and customers will be very happy, they will pay AMD back in the future with more purchases

Won't happen. Because AMD needs to make money for further R&R and because of this:

In addition too low of a price right now sets up the wrong expectations about future pricing going forward. 3nm chips are very expensive in terms of R and D and cost of wafers.

Exactly. If they go too low now, people will inevitable complain when prices go up again and hold of purchasing any new card.

 

[gaav87]

For ppl who are saying its not ready some polish reviewers already saw them working in november in taiwan.

 

[Timorous]

So a 7900 XTX and 4080 are 1440p cards as well?

You could make that argument given the frequent use of FSR / DLSS to hit 4K at decent frame rates.

 

[Meteor Late]

Exactly. If they go too low now, people will inevitable complain when prices go up again and hold of purchasing any new card.

Yeah, which is why they shouldn't raise prices and that's it.

 

[branch_suggestion]

So delaying until March is looking better by the minute, all the people who got mad for selling their current cards based on rumours and not listening to AMD (who only said Q1), wouldn't be able to buy Blackwell anyway.
So either you wait for Blackwell to be in decent volumes by April, or you buy the RDNA4 parts that have been ready for a few months longer with actual supply.
Bad news is that means pricing is likely to be higher.

 

[adroc_thurston]

Bad news is that means pricing is likely to be higher.

Bad news is AMD killing top of the stack with no replacement when the comp completely shits itself.

 

[branch_suggestion]

Bad news is AMD killing top of the stack with no replacement when the comp completely shits itself.

Yeah, no R600 successor until Hawaii does that.
384-bit Cypress, that is all AMD had to do and just go from there.
What was the point of Eyefinity without the performance crown?

 

[adroc_thurston]

What was the point of Eyefinity without the performance crown?

It was a cool display core gimmick I guess.
But AMD is the only company not Apple still making good display engines so nothing changed.

 

[branch_suggestion]

But AMD is the only company not Apple still making good display engines so nothing changed.

Navi10 was a big oopsie though.

 

[adroc_thurston]

Navi10 was a big oopsie though.

Fixed in Navi14 so no biggie.

 

[lixlax]

Could one of the main reasons for the weird(ish) march launch be that AMD and its partners are right now trying to stockpile as many cards as possible in the US before the tariffs hit? So come march they'll have enough manufactured that rest of the world can have a decent number as well and won't be complaining of it being a paper launch.

The other reason being, as discussed, launching the cards after NV 5070 series so they couldn't correct the price before launch. I'm fairly sure that AMD has very good knowledge of 5000 series' performance etc characteristics at this point.

 

[Meteor Late]

Cancelling Navi 41 could end up being Lisa Su's biggest blunder as AMD CEO.
From what we know about the expected performance of RTX 5000 lineup and 9070 XT, along with die size info, it seems RDNA4 has similar or close enough PPA in the same node against Nvidia, using considerable worse memory. This has to be the first time this happens since god knows when. Also, finally good enough Raytracing and a good upscaler in FSR4.
RDNA4 was the Zen moment, it is a HUGE wasted opportunity.

 

[del42sa]

Cancelling Navi 41 could end up being Lisa Su's biggest blunder as AMD CEO.
From what we know about the expected performance of RTX 5000 lineup and 9070 XT, along with die size info, it seems RDNA4 has similar or close enough PPA in the same node against Nvidia, using considerable worse memory. This has to be the first time this happens since god knows when. Also, finally good enough Raytracing and a good upscaler in FSR4.
RDNA4 was the Zen moment, it is a HUGE wasted opportunity.

I am sure, there was a good reason, either high price or limited functionality or both

 

[ToTTenTranz]

For ppl who are saying its not ready some polish reviewers already saw them working in november in taiwan.

All the git pushes on linux distros tell the same story about software being mostly ready since Summer, by the way.


6 months sitting on perfectly working and high performant hardware before release. 🤦‍♂️

 

[Heartbreaker]

Cancelling Navi 41 could end up being Lisa Su's biggest blunder as AMD CEO.
From what we know about the expected performance of RTX 5000 lineup and 9070 XT, along with die size info, it seems RDNA4 has similar or close enough PPA in the same node against Nvidia, using considerable worse memory. This has to be the first time this happens since god knows when. Also, finally good enough Raytracing and a good upscaler in FSR4.
RDNA4 was the Zen moment, it is a HUGE wasted opportunity.

Unless there is new Die Size info, the last I have seen Navi48 is sitting about 390mm^2 vs 378mm^2 for GB 203, and NVidia step down version (3070 Ti) is expected to outperform AMDs Top (9070 XT) version, that gives NVidia as significant PPA lead.

So it would essentially be suicide for AMD to chase 5090.

So this looks more like a smart move than a blunder.

 

[insertcarehere]

Cancelling Navi 41 could end up being Lisa Su's biggest blunder as AMD CEO.
From what we know about the expected performance of RTX 5000 lineup and 9070 XT, along with die size info, it seems RDNA4 has similar or close enough PPA in the same node against Nvidia, using considerable worse memory. This has to be the first time this happens since god knows when. Also, finally good enough Raytracing and a good upscaler in FSR4.
RDNA4 was the Zen moment, it is a HUGE wasted opportunity.

Lisa's has plenty of blunders under her tenure, canceling a niche high-priced consumer part for an inherently niche segment will almost certainly not be one of them..

 

[Meteor Late]

Unless there is new Die Size info, the last I have seen Navi48 is sitting about 390mm^2 vs 378mm^2 for GB 203, and NVidia step down version (3070 Ti) is expected to outperform AMDs Top (9070 XT) version, that gives NVidia as significant PPA lead.

So it would essentially be suicide for AMD to chase 5090.

So this looks more like a smart move than a blunder.

I've seen other figures less than 390 mm2.
Either way, 5080 won't be that much faster than 5070 ti, which I expect 9070 XT to match, and Nvidia is enjoying like 40% higher memory bandwidth or so.

 

[Timorous]

Unless there is new Die Size info, the last I have seen Navi48 is sitting about 390mm^2 vs 378mm^2 for GB 203, and NVidia step down version (3070 Ti) is expected to outperform AMDs Top (9070 XT) version, that gives NVidia as significant PPA lead.

So it would essentially be suicide for AMD to chase 5090.

So this looks more like a smart move than a blunder.

I don't see how 390mm can possibly be true for 64CUs and 256bit bus.

That is bigger than N32 + the MCDs and besides the MCDs being on N6 there is also the fact both the GCD and MCD contain the links to talk to each other which is die area that does not need to be used in a monolithic design.

To me the realistic upper bound is probably in the 320mm region assuming the new RT features and other changes make the CUs larger than in RDNA3.

 

[GTracing]

Intel losing 3.5 billion on ARC(this figure was in 2022), just shows how expensive GPU development is.

Intel didn't lose 3.5 billion on Arc. That's one firm's estimated total loses for Intel's AXG division. AXG included desktop dGPUs, datacenter dGPUs, server AI accelerators, and HBM server CPUs.

 

[GTracing]

I don't see how 390mm can possibly be true for 64CUs and 256bit bus.

That is bigger than N32 + the MCDs and besides the MCDs being on N6 there is also the fact both the GCD and MCD contain the links to talk to each other which is die area that does not need to be used in a monolithic design.

To me the realistic upper bound is probably in the 320mm region assuming the new RT features and other changes make the CUs larger than in RDNA3.

Yeah, the numbers don't add up. Either the 390mm is wrong. Or N48 has a boatload of cache. Or it's on N6. Or the compute units are substantially bigger.

 

[Win2012R2]

Or the compute units are substantially bigger

They are expected to be bigger since they finally had to add proper stuff for RT and their equivalent of tensors, plus big cache most likely to deal with slow GDDR6

 

[insertcarehere]

Yeah, the numbers don't add up. Either the 390mm is wrong. Or N48 has a boatload of cache. Or it's on N6. Or the compute units are substantially bigger.

The 390mm^2 figure came from a Twitter user comparing pics of 'N48' die to Zen packaging side-by-side. FWIW the methodology looks good enough that IMO there won't be significant deviation from that figure.

 

[ToTTenTranz]

They are expected to be bigger since they finally had to add proper stuff for RT and their equivalent of tensors, plus big cache most likely to deal with slow GDDR6

There are no tensor units in RDNA4. Tensor ops still run on the shader units.

 

[Win2012R2]

There are no tensor units in RDNA4

Still???

What are they going to be running FSR4 on then - shaders again?