Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

Z O X · Jan 9, 2025

insertcarehere said:
I'd be very surprised that the AIBs would tolerate $549 with AMD in the first place. Looking at the furmark leak:

The GPU can max at 330w TBP with GPU temps in the mid-50s while the fan is at 1100rpm or so.. Correct me if I am wrong but any cooler that can do this are gonna involve very substantial metal heatsinks, and good quality fans (read: $$$), not to mention power delivery that can handle 3x 8-pin. These costs are all borne by the AIB who would not be happy if AMD's guidance price is that low (ballpark to similar designs for 4080S/7900XTX?).

Having said that, with these performance indications even $599/$649 should get good market response..

VRAM 10C over hotspot?
Anyway, my 2,5 slot RX6800XT hovers around 90-95C on hotspot at similar wattage.
This looks like a water cooled product, but memory temp worries me ...

SolidQ · Jan 9, 2025

Z O X said:
but memory temp worries me ...

That AIB design, other AIB can be better

Area Efficiency

https://twitter.com/x/status/1877287284606701909

jpiniero · Jan 9, 2025

^^^

Average 4090 gets about 19k, so 24k is about 22% faster for 33% more SMs. What do you think is going to happen with the 5080 and 5070Ti, which only have 5% more?

It's possible the 9070 XT is faster than the 5080 in raster.

insertcarehere · Jan 9, 2025

gaav87 said:
u are arguing about 1 pizza price for next 2years of use...

The AIB business is not exactly something with lots of margins to begin with so believe it or not that Pizza slice (which I doubt good vapor chambers are) matters. AMD with its current market position has very little leverage to tell these guys to just take it on the chin.

Keller_TT said:
I think AIB cards for the 9070 XT would be up to $600. They can reuse the designs from overclocked 7900 GRE with updated materials and thermals in line for 2025 .

Looks can be deceiving.. But many of the cards we're seeing sure don't look like they could've been 'overclocked 7900 GRE' designs, or at least ones that weren't reused wholesale from 7900XTX/XT in the first place:

Keller_TT · Jan 9, 2025

insertcarehere said:
The AIB business is not exactly something with lots of margins to begin with so believe it or not that Pizza slice (which I doubt good vapor chambers are) matters. AMD with its current market position has very little leverage to tell these guys to just take it on the chin.

Looks can be deceiving.. But many of the cards we're seeing sure don't look like they could've been 'overclocked 7900 GRE' designs, or at least ones that weren't reused wholesale from 7900XTX/XT in the first place:

View attachment 114579
View attachment 114580 View attachment 114581
View attachment 114582

From what I read, the ASUS Prime is a 2.5 slot design. Far from 7900 XTX. The beefier TUF is also within 3 slots iirc.
I'm not so sure that the final drivers will push TDP beyond 310W for that final 2% and loose the efficiency value as well as market positioning wr.t 7900 XT. The XTX with 8 more Gigs VRAM is meant to be a tier above with how AMD is positioning RDNA 4 in name and price.

eek2121 · Jan 9, 2025

insertcarehere said:
I'd be very surprised that the AIBs would tolerate $549 with AMD in the first place. Looking at the furmark leak:
View attachment 114572
The GPU can max at 330w TBP with GPU temps in the mid-50s while the fan is at 1100rpm or so.. Correct me if I am wrong but any cooler that can do this are gonna involve very substantial metal heatsinks, and good quality fans (read: $$$), not to mention power delivery that can handle 3x 8-pin. These costs are all borne by the AIB who would not be happy if AMD's guidance price is that low (ballpark to similar designs for 4080S/7900XTX?).

Having said that, with these performance indications even $599/$649 should get good market response..

Probably cooler reuse from past generations or from NV RTX 50. Besides, you are talking a few dollars difference tops. Heatsinks can be made dirt cheap these days. Automation helps tremendously.

marees · Jan 9, 2025

"One of the weirder GPU launches"

https://twitter.com/x/status/1877359933659951371

poke01 · Jan 9, 2025

poke01 said:
I’m breaking my rule. If these cards are real deal, I might get one. Cause why not….. I love GPUs 😭😭😭

After waking up, I think it’s best to wait for DLSS4 improvements and keep my card for one more generation. Feels like a waste to upgrade so soon since this 4070S has zero issues.

For those that are upgrading enjoy RDNA4, looks to be a great architecture from these time spy bench’s, hopefully it translates to games too.

carrotmania · Jan 9, 2025

Industry leading PPA for GPU, chip is known to be half of what they planning, software finally catching up to NV, FSR4 looks great, highest performing CPU to market it with, cards from multiple AIBs on the trade floor...

AMD be all like "we have a plan to tank our share price!"...

If this delay isn't because Lisa took a swan dive into a bin in the RnD room to try and find BigNavi to show us next week... I'mma be very pis... annoyed...

Edit: I see Ryan Smith said much of the same thing, yeesh

adroc_thurston · Jan 9, 2025

marees said:
"One of the weirder GPU launches"

they didn't even try shilling FSR4 and it looks good™.
You'd think they cart shill foundry out. alas.

lucasworais · Jan 9, 2025

We can dream, right?

carrotmania said:
If this delay isn't because Lisa took a swan dive into a bin in the RnD room to try and find BigNavi to show us next week... I'mma be very pis... annoyed...

exquisitechar · Jan 9, 2025

adroc_thurston said:
You'd think they cart shill foundry out. alas.

They really ought to get on that + respond to NV’s frame generation update ASAP or it won’t matter at all if RDNA4 is actually good in terms of PPA. From what I’m seeing, the average person is mesmerized by Nvidia’s exclusive, must-have “features” more than ever before.

jpiniero · Jan 9, 2025

exquisitechar said:
They really ought to get on that + respond to NV’s frame generation update ASAP or it won’t matter at all if RDNA4 is actually good in terms of PPA. From what I’m seeing, the average person is mesmerized by Nvidia’s exclusive, must-have “features” more than ever before.

Well, if NV can rip off Lossless Scaling, AMD can too.

Hitman928 · Jan 9, 2025

carrotmania said:
Industry leading PPA for GPU, chip is known to be half of what they planning, software finally catching up to NV, FSR4 looks great, highest performing CPU to market it with, cards from multiple AIBs on the trade floor...

AMD be all like "we have a plan to tank our share price!"...

If this delay isn't because Lisa took a swan dive into a bin in the RnD room to try and find BigNavi to show us next week... I'mma be very pis... annoyed...

Edit: I see Ryan Smith said much of the same thing, yeesh

If Lisa was the one to kill it (I doubt it but maybe) it's only because the Radeon team did a poor job of selling it to her. If they had come to her and said we can make a chip that matches or beats NV's best and do it at less cost and have feature parity, I don't see how she would say don't do it. Instead, they were probably doubtful themselves about their competitiveness given what NV pulled out the last generation and so why would you pour money into a product your people are saying is doomed to fail?

CastleBravo · Jan 9, 2025

marees said:
"One of the weirder GPU launches"

https://twitter.com/x/status/1877359933659951371

Nvidia put AMD in a hard spot when their presentation basically said nothing about the 5070's performance other than "stronger than 4090 with multi frame generation", and with a pricetag less than the 4070's launch MSRP. This was a brilliant move by Nvidia as AMD can't make a similar claim of 4090-class performance, and they were probably expecting the 5070 to be priced higher than the 4070 at launch, not less. Any performance numbers put out by AMD will look weak in the eyes of many less informed consumers at least until all the youtube reviewers show everyone exactly how weak the 5070 is compared to the 4090. I suspect AMD's driver team is working mega overtime right now to rush out their own version of a Lossless Scaling 4x framegen rip off.

Win2012R2 · Jan 9, 2025

CastleBravo said:
AMD can't make a similar claim of 4090-class performance

Yes they can, but most people will angrily say they used fake frames to achieve it, rather than using DLSS 4 that can do no wrong...

Hitman928 said:
I don't see how she would say don't do it

I can easily see it - she just say: we have to do MI300 instead and make 12-15x the money you can possibly hope to make with big consumer chip, she isn't not wrong sadly...

carrotmania · Jan 9, 2025

Hitman928 said:
If Lisa was the one to kill it (I doubt it but maybe) it's only because the Radeon team did a poor job of selling it to her. If they had come to her and said we can make a chip that matches or beats NV's best and do it at less cost and have feature parity, I don't see how she would say don't do it. Instead, they were probably doubtful themselves about their competitiveness given what NV pulled out the last generation and so why would you pour money into a product your people are saying is doomed to fail?

But, that was assuming that NV would pull off another 4090, and they accidentally another 2080 instead. So lets hope Dave from RnD just put it quietly in a drawer somewhere, and the reason Lisa didnt do the keynote is cos she was down on her knees helping them look for a 2x9070XT which sounds like it would walk all over the 5090...

igor_kavinski · Jan 9, 2025

Heartbreaker said:
If they force NVidia to reprice they have gone too far.

That might be an unpopular opinion, but it isn't AMDs job to create a price war for our benefit, since they still need a healthy profit/part.

And a price war won't necessarily gain them market share. It they goad NVidia into lower prices, they end up back in the same situation, with the same price gap, but now with less profit.

Oh wow. Your concern for AMD's financial health is duly noted. Do I smell fear?

Nvidia can burn if AMD's lower prices help do that. AMD will make Nvidia look bad if their price/performance ratio is excellent and AMD has plenty of FU money from their enterprise sales. Making a competitor weaker is par for the course and AMD will definitely be justified in doing that since Nvidia has never pulled their punches in screwing over AMD with their "free" online marketing teams of neanderthals that send their burned GPUs for repairs rather than disowning Nvidia.

Keller_TT · Jan 9, 2025

Since RDNA4 looks like a good one, and so does Strix Halo, here is some interesting news looking ahead.
Rapidus, the Japanese Fab in the works with Sony as a founding partner and investor, 2nm process co-developed with IBM, and Jim Keller as advisor+customer, is wooing TSMC's 2nm clientele.
I think the PS6 custom SoC will be Made in Japan™ for 2027. A topic I'll keep an eye on as we move on. I wish them all success.

gdansk · Jan 9, 2025

igor_kavinski said:
AMD has plenty of FU money

No, Nvidia has FU money. How do you weaken a competitor that has $38 billion in cash on hand?

gaav87 · Jan 9, 2025

Or maybe AMD just has to start releasing good products and this is another x3d moment ?
96CU rDNA4 24GB would stomp on 4090 and 5080.

adroc_thurston · Jan 9, 2025

Keller_TT said:
I think the PS6 custom SoC will be Made in Japan™ for 2027

Lmao no, Rapidus is whatever at zero volume.

gaav87 said:
96CU rDNA4 24GB would stomp on 4090 and 5080.

Yeah dawg that's the plan for RDNA5. They insert one more die at the kilobuck price point.

igor_kavinski · Jan 9, 2025

gdansk said:
No, Nvidia has FU money. How do you weaken a competitor that has $38 billion in cash on hand?

Maybe not financially but in mindshare. Sooner or later, even diehard zombies will wake up and realize that they are being taken for fools by Jensen.

del42sa · Jan 9, 2025

adroc_thurston said:
Yeah dawg that's the plan for RDNA5. They insert one more die at the kilobuck price point.

I'll believe it when I see it ...

Keller_TT · Jan 9, 2025

adroc_thurston said:
Lmao no, Rapidus is whatever at zero volume.

Why? that's as of now. The news from there is that they're building steadily and will start sampling this year and build up for HVM by Q4 next year, and see no major hurdles for high yields going into 2027.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Junior Member

Golden Member

Lifer

Senior member

Member

Diamond Member

Senior member

Diamond Member

Member

Diamond Member

Member

Senior member

Lifer

Diamond Member

Member

Senior member

Member

Lifer

Member

Diamond Member

Senior member

Diamond Member

Lifer

Member

Member