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

 

[uzzi38]

Average gamer buys whatever has the brand rep and you could only build the brand rep by winning.

Again, the Intel example shows otherwise.

no one cares about Arc outside of "their GFX IP will be good enough to force NV pricecuts eventually. Be. Serious.

You're just proving you have no clue what people think now, to be frank.

one NV pricecut away from irrelevancy.

YESSSSSSSSSSSSSSSSSSS.
when you're behind, you have to gamble! Stake yourself on being bold, and ambitious and ship something you think comp can't. Something that pulls a decisive win on top.
That's how AMD won server CPUs ffs. And client CPUs, really.
That's where Radeon guys have been failing since the goddamn R420.

That's not a gamble, that's a death knell.

Going to say it again, there's no point to chasing halo without a few generations of "consistently decent" before that. You have to give people confidence you're worth a damn before you can convince them you're the best.

Another example here: look at Strix Halo. Reviewed amazingly, there are people recognise how good it is as a mobile product.

Very few people are actually excited for it though, as they see it as a further extension of "Ryzen mobile has no supply and is always overpriced, and will forever remain that way". It doesn't matter how much you point out stuff like pricing of the Z13, Strix Halo is a pretty boring product to the masses because AMD has built up a reputation in mobile of just not being able to supply the market, and even when they can, they're just too expensive.

 

[adroc_thurston]

Again, the Intel example shows otherwise.

It really doesn't.

You're just proving you have no clue what people think now, to be frank.

They don't think. They buy Nvidia.

That's not a gamble, that's a death knell.

No that's a gamble. Anything you do in semis is a gamble, really. Some are bigger than others.
When you're behind, you make a bigger, bolder bet on wackier things.

Going to say it again, there's no point to chasing halo without a few generations of "consistently decent" before that.

R300.

You have to give people confidence you're worth a damn before you can convince them you're the best.

R300.

Another example here: look at Strix Halo. Reviewed amazingly, there are people recognise how good it is as a mobile product.

not R300.

Very few people are actually excited for it though, as they see it as a further extension of "Ryzen mobile has no supply and is always overpriced, and will forever remain that way". It doesn't matter how much you point out stuff like pricing of the Z13, Strix Halo is a pretty boring product to the masses because AMD has built up a reputation in mobile of just not being able to supply the market, and even when they can, they're just too expensive.

not R300.

 

[Timorous]

AMD can pull that off without resorting to the lengths Intel has. They don't need to ship a die twice as large with more VRAM to provide compelling value at an even lower price. I don't think they should be selling 9070XTs for $550. But I don't see how $599-$629 wouldn't be both a huge improvement for Radeon over last gen, whilst simultaneously being a huge step up in popularity.

AMD are in a no win situation. The market is quite hungry so if 9070XT came in at $600 it will sell out pretty fast unless AMD have really robbed peter on wafer supply, which I doubt. Then when it sells out and is scalped people will moan about low supply and paper launch and on and on as a way to convince themselves spending $900 on a 5070Ti is not that bad really given how there is no other competition.

Best to talk about what features could potentially lock someone out of buying Radeon after we've seen FSR4, and some other stuff that will become clear once the launch is done. All I'll say for now is AMD is attempting to break down that software moat one bit at a time.

It will just move to ray reconstruction or g-sync or back to drivers or CUDA or some other NV feature that AMD don't match them on.

Again, the Intel example shows otherwise.

I don't think the people here, on reddit, even places like GAF are what you would call average gamers. They are far more informed than the average gamer.

 

[Jan Olšan]

Intel's a clear example - if a very extreme one - that price is what it takes to win over gamers. Nobody cares about driver issues or anything else, B580 is seen as a clear winner product for a simple reason: 4060Ti performance at below 4060 pricing.

Intel is probably an extreme. And they may also find that now they have a "hey you suck you aren't 30% cheaper for the same performance" entrenched thinking to deal with.

The truth is likely somewhere in the middle, but I would like to remind everyone that the only thing that can change AMD marketshare in GPUs is OEM (full PC) sales. Retail is small minority of the market so you can't move the overall indicators much even if you totally own it. See Intel processors after all. (and notebooks, good luck agianst those OEM deals...)

Actually, people keep repeating the has to be cheap × gotta make money arguments all over, but we should first stop and think about what is the retail marketshare they got now. If they had 12 % there, then yeah, would be room to grow (and gain 2% in overall gDGPU JPR number?). But what if people actually already do buy Radeons in retail, hmm? Something that really needs analyzing before we start with "AMD MUST DO THIS" steve videos.

 

[adroc_thurston]

Retail is small minority of the market

Not really, no, standalone AIC sales are pretty huge.

 

[exquisitechar]

Well the reality is, they did, and they have to live with a new market now that looks at Intel more favourably than they do at AMD.

Making yourself look favorable by pricing your GPUs into oblivion isn't what AMD wants, presumably. Of course, they could go for a far less extreme version of that and offer compelling value at $599 while still making okay money, as you said, but I think Timorous is right that that would barely make a difference in the long term.

At this point, I agree that the only way towards a real turnaround is AMD leveraging their expertise to create a 3D stacked halo part that Nvidia can't match with their giant monolithic GPUs. That alone would be a huge deal, since no one can stay blind to Nvidia's fastest not being number one in most benchmarks. If that works out, they then need to keep being consistent. If they stumble once, they can't panic and release a half-assed lineup the next generation like they did with RDNA4. Never completely ceding the high end market without even showing up for a fight. The fact that 7900XTX owners have nothing but a sidegrade to look forward to from AMD makes them look bad. I see casual gamers look at the 9070XT and think that AMD is washed up and can't touch Nvidia, since they don't know or care about how much better RDNA4 is than RDNA3. Just making a regular, monolithic, 384-bit part that beats the 5080 would have been a lot better than what they actually did. Simply put, there's no easy way to turn things around when Nvidia dominates as much as it does. AMD needs to be ambitious and seriously invest in their client dGPUs, consistently, or accept their growing irrelevancy and potential slide into an exit of the market altogether.

 

[adroc_thurston]

AMD needs to be ambitious and seriously invest in their client dGPUs, consistently,

They're already throwing money at rebuilding their shader cores from scratch each gen.
What they need is balls.
They can ask Forrest Norrod to loan his idk. He won, and they didn't.

 

[Jan Olšan]

At this point, I agree that the only way towards a real turnaround is AMD leveraging their expertise to create a 3D stacked halo part that Nvidia can't match with their giant monolithic GPUs. That alone would be a huge deal, since no one can stay blind to Nvidia's fastest not being number one in most benchmarks.

I wonder how would the goalposts move. Micorbenchmarks showing that the architecture is inferior even if performance is superior? Good old cuda/dlss/"just works"?
I think I recall such stuff from early Threadripper times.

 

[eek2121]

Meh, even with this gen, a 96CU-128CU halo part would have done the job IMO. Simply using N3 would have also given them a decent boost as well.

 

[gaav87]

They don't, NV has a comfortable margin cushion to force a price war if needed.
N48/44 are worthless products (waste of a really, really cool microarch, unfortunately).
Like any client gfx lineup without a halo part is worthless.

You forgot one thing. A LOT of people are on hd7870 rx580 gtx1660 or just straight up gave up on pc gaming just waiting in shadows.
449/499$ 4080s would make all of them move. Nvidia buyers already got their burnwell at 1-2k$ and dont even care about the missing rops on reddit.
If they drop this bomb it will sell look at 9800x3d. They priced it competitive and it was out of stock for months even at higher prices same thing would happen here.
It would sell so good that they would not be able to keep up with the production for 3-4 months.
They would win hearts and minds of gamers. All <redacted> would hype this up as next comming of jesus and mock every 5070 5070ti 5080 user.

Post edited, the use of "fanboys" or similar derogatory words for vendor fans is not permitted in tech. -Moderator Shmee

 

[tsamolotoff]

They would win hearts and minds of gamers

First they need to win the minds of the tech press, which is throughly astroturfed and many of them like GN-Steve even welcome it. As some other people mentioned, thanks to software dominance, nV will invent another metric (neural shaders, some other AI bs) in which any Radeon offering will be a loser by default.

The last true failed halo from AMD was Fiji

It wasn't even halo, it was an emergency SKU that was shipped because 20nm was stillborn. Compounded with the fact that AMD was at the bottom financially, i'd say it was the moment when they've sacrificed Radeon for Ryzen.

after we've seen FSR4

nV foundry will post dozens of videos of how FSR4 is objectively bad if you magnify the picture by 500% and it is naturallly inferior to DLSS4 in carefully curated sequence of still shots (as I said earlier, I'm against all forms of temporal reconstruction so this doesn't matter for me, but the people who don't mind blur will believe this to be worth 20% markup etc)

The only time Radeon wasn't declining was R300 or them coasting off R300 goodwill.

What's wrong with R580? I'd say it's just big failures like R600 or NV40 that leave lasting impressions, while successes are not really that sticking with regards of the sentiments of the public (we'll see how the Raptorlakegate will affect Intel dcpu sales)

 

[tsamolotoff]

Sure, and still can't say I'm very impressed

Interesting, I think I've got like 11.5k on my not so fresh Debian testing installation, but I couldn't properly test it as I've failed to replace pptable, it just locks up if I replace it with upp, or plainly refuses to get replaced via plain copy

 

[Gideon]

Either way, it's pretty obvious where it's at

$550 / $650 would be a reasonable compromise (in margins, reviewer ridicule, etc). This is something they should do. Perhaps the XT could be 20-50$ cheaper (if they want it to be an upsell)

Unfortunately, $650 / $700 looks to be the thing they are actually doing (based on leaked pricing) and that will suck as:

• Nvidia -50$ is not a good idea for optics (even when the price is fictional) - that's only a 7.5% price difference (for slightly worse RT perf and less features)
• 50$ between the XT and non-xt is way too small, it should be at least 70$ (even for an upsell)
• Reddit will whine anyway ( in their "it should be $400" bubble) but this will cause all of the reviewers to ridicule it too

All the points about shifting goalposts in features, DF DLSS4 shilling and Nvidia correcting prices are valid, but miss the big picture. AMD doesn't have to convince 50% of the market, they only need to convince some (already doubling their share back to 20% would be extremely good)

Also Nvidia is heavily supply limited for a while. they can't really change the prices for at least a few months.

Therefore, the cheaper price would be worth it for the positive reviews alone. Yes there will be claims of "paper launch", etc, but that's water under the bridge. If it's 700 AMD 100% has to drop the price in a few months (as they always do) ...

Besides, with a revenue like that, Nvidia might not even care and let AMD claw back some marketshare:

 

[gaav87]

First they need to win the minds of the tech press, which is throughly astroturfed and many of them like GN-Steve even welcome it. As some other people mentioned, thanks to software dominance, nV will invent another metric (neural shaders, some other AI bs) in which any Radeon offering will be a loser by default.

It wasn't even halo, it was an emergency SKU that was shipped because 20nm was stillborn. Compounded with the fact that AMD was at the bottom financially, i'd say it was the moment when they've sacrificed Radeon for Ryzen.


nV foundry will post dozens of videos of how FSR4 is objectively bad if you magnify the picture by 500% and it is naturallly inferior to DLSS4 in carefully curated sequence of still shots (as I said earlier, I'm against all forms of temporal reconstruction so this doesn't matter for me, but the people who don't mind blur will believe this to be worth 20% markup etc)


What's wrong with R580? I'd say it's just big failures like R600 or NV40 that leave lasting impressions, while successes are not really that sticking with regards of the sentiments of the public (we'll see how the Raptorlakegate will affect Intel dcpu sales)

LOOK at the size of nvidia gaming revenue... I don't think they would even move the prices by 50$ IF AMD did the move i proposed. Jensen would not even flinch, a muscle,
for 300-400milion$ he would just allocate more wafers to AI. So the price war adroc is talking about would not even happen.
But all the waiting crowd on old rx580 rx5700 gtx1660 gtx1070 gtx1080 even rtx3060 would hop on the AMD train. Everyone would want 9070 9070xt you guys and amd underestimate the old gpu crowd and focus on guys who upgrade every gen. But there are ppl who upgrade every 2 gens like me or every 4 or even 6 gens and they are right now up for grabs.
All the big dogs would still buy 5070ti/5080/5090 anyway like they do right now

 

[SolidQ]

Hellhound have small OC and Navu ofc

 

[Gideon]

Navu

That sounds like something Yeston would release "Navu Waifu edition" or sth

 

[gaav87]

Hellhound have small OC and Navu ofc

Word on street is rdna4 has huge supply.

 

[branch_suggestion]

Real sentiment is Twitch chat et al.
But nobody has the patience to start GPU flamewars there just to test a theory.
But I've seen enough to know the sentiment there is extremely NV leaning, which seems to reflect the market most accurately.

 

[SolidQ]

Real sentiment is Twitch chat et al.

Yreah i saw people force NV, a lot people even doesn't know about Radeon and still a lot people thinking Radeon drivers are bad(those people seems still sitting in medieval ages)

 

[ToTTenTranz]

The proof is in historic mss data.
The only time Radeon wasn't declining was R300 or them coasting off R300 goodwill.
End of the story.

That was 23 (twenty-three) years ago.
Back when we had major new architectures every year or year and a half, when 150nm waffers cost $1000 or lower, when UMC was still in the race for end processes, internet wasn't a global thing, we had no smartphones just dumb mobile phones, Google was a 4 year-old startup, GPUs were 1000x smaller than they are today and had 1000x less VRAM.

Making the whole basis of your argument around R300 is bonkers, honestly.



AMD did gain a truckload of marketshare with Polaris, they went from 20% to 30% in 2 years, and despite the lacking features RDNA1 actually managed to withhold that marketshare a bit.

It was their Nvidia -$50 strategy of RDNA2 and RDNA3 between 2020 and 2024 that killed the mindshare among PC users, and the brand reputation has never been worse than it is right now.

Average gamer buys whatever has the brand rep and you could only build the brand rep by winning.

Another thing you didn't have 23 years ago: youtube reviewers/influencers.
If they get LTT + GamersNexus + HWUB + J2C + etc. rooting for those cards, there's the brand rep you talk about.

It's one thing for those guys to say "this card is crap, all this generation is crap" while not being able to provide an alternative. If they say "come here and buy this because it's good" and there's enough availability, the scenario is completely different.

one NV pricecut away from irrelevancy.

Geforce prices are sky high because they're allocating most of their wafers to the AI and top-end chips that get them ridiculous margins.

If Nvidia cuts the price on the 5070-5080 with their current volume, AIBs and distributors will simply take a bigger piece of the pie by scalping the customers even more. The only way for Nvidia to be able to reduce the price is if they start allocating more wafers to GB203 and GB205 instead of GB202 and GB100.

And they would be crazy to do that with the current demand for the higher end chips. Geforce amounts to 10% of their current revenue.

 

[Tuna-Fish]

They don't but when they cannot supply enough 9800X3D to keep it in stock and that on its own sells for $500 how do AMD justify a $500 9070XT when they are supply limited in another product that has far fatter margins?

The two products do not compete for resources. X3D is still limited by stacking capacity, not silicon.

 

[Timorous]

It was their Nvidia -$50 strategy of RDNA2 and RDNA3 between 2020 and 2024 that killed the mindshare among PC users, and the brand reputation has never been worse than it is right now.

I disagree on this.

It was DLSS actually coming good (and being pushed) as well as RT to a smaller degree that cemented NV in their current position.

Arguably the 6800XT did have a killer feature over the 3080, 60% more VRAM. The thing is though that got discounted because DLSS / RT was all that was pushed and devs went in that direction. Maybe AMD should have spent more time with devs developing high res texture packs that would exceed the 10GB 3080 frame buffer but run happily on the 16GB (or 12GB 6700XT) buffer while providing a decent increase in IQ and texture variance.

Of course that is where the double standard comes in. When AMD have a feature advantage it gets downplayed a lot and when NV have a feature advantage it becomes the be all and end all. Of course the fact most in game TAA algorithms are utter trash helps the situation for NV (pure conspiracy theory but maybe NV encourage their partners to have rubbish in game TAA to make DLSS look even better comparatively) since the DLSS TAA algorithm is a lot better.

 

[QuickyDuck]

Halo won't saves AMD really.
It needs AI sales to justify the cost and we know AI people aren't getting these cards unless they have well supported AI software stacks. Gamers won't pick it up either. All they want is cheaper NV card. Halo gonna be a niche in AMD fans hand so it's basic just a money hole.

What AMD need is muti-gen of good products and builds their reputation.

 

[ToTTenTranz]

I disagree on this.

It was DLSS actually coming good (and being pushed) as well as RT to a smaller degree that cemented NV in their current position.

The Nvidia -$50 strategy failed precisely because people acknowledged that Nvidia's exclusive features are worth more than $50. In 2020 and 2022, those were DLSS and (to a lesser degree) RT performance.

 

[Glo.]

yeah they should, that's their last chance before Lisa cuts dGFX guys altogether.

No. Just, no.