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

 

[marees]

I am a bit confused here, as it seems you may be saying contradicting things, at least to me. You say that nothing will replace RDNA 3 at the high end, but then say that an 240mm die will be the high end. Assuming this new die is RDNA 4, either it is faster and it replaces the 7900XTX, or it is slower and it doesn't. It may be newer, but if it isn't faster, then it isn't the new top end card.

I suppose I should clarify that by high end, I meant 7900XT and above, mainly thinking the XTX. Still, if the 7900XTX remains the fastest AMD card for some time, then it isn't replaced, so long as there is stock, new or used. I suppose they could stop producing them, and supply could start to dry up, but that doesn't make the 7900XTX any slower. Replacing it would seem a very bad move, if there isn't anything faster.

It is so weird. AMD had navi 36 & also a 3d v cache of N31 & also RDNA 3.5 APUs, but no sign of replacement for the 'flawed' N31 parts. I don't get it at all. Hopefully AMD will tease a high level roadmap in June at computex

 

[igor_kavinski]

AMD is doing the best they can with what they have. Going against established (and manipulative) companies like Intel/Nvidia is not something they can afford to do with an outright price war because both of those companies can fire those shots wayyyy better until AMD has no choice but to bow out of the market. It's not fair to the little guy buying crap from Intel (their space heaters) and Nvidia (proudly and shamelessly selling an overpriced product with reliability issues) but then the little guy refuses to do his research and just subscribes to the herd mentality without question. I don't see things changing any time soon.

AMD can't please anyone in their situation.

If they sell faster stuff at lower prices, they won't be able to handle the volume and people will still buy Intel/Nvidia because those two will come up with clever/deceptive marketing tactics to sell their crap.

If they sell their faster stuff at Intel/Nvidia prices, well, then the consumers won't be happy about that and most of them may still go with Intel/Nvidia.

I just don't see any win-win situation for AMD, other than the enterprise sector where they have built up great loyalty through their power efficiency because every megawatt there counts.

 

[igor_kavinski]

Something really must be inbound. Prices dropping in anticipation and to clear existing inventory: https://videocardz.com/newz/amd-radeon-rx-7700-xt-price-has-just-hit-379

 

[H T C]

AMD can't please anyone in their situation.

If they sell faster stuff at lower prices, they won't be able to handle the volume and people will still buy Intel/Nvidia because those two will come up with clever/deceptive marketing tactics to sell their crap.

If they sell their faster stuff at Intel/Nvidia prices, well, then the consumers won't be happy about that and most of them may still go with Intel/Nvidia.

I just don't see any win-win situation for AMD, other than the enterprise sector where they have built up great loyalty through their power efficiency because every megawatt there counts.

Slightly more performance for a similar price or a slightly lower price for the same performance WILL NOT BE ENOUGH.

Correct me if i'm wrong but, has AMD EVER sell a graphics card, and i'm referring since launch day, like they did ZEN 1st generation relative to Intel's then offerings?

They undercut Intel MASSIVELY: either vast more performance @ similar price or vastly lower price for similar performance: THAT'S what they need to do in the GPU scene.

 

[MoogleW]

Hm did not know that rdna4 aimed to simplify the wgp. Maybe its 8x simds that are 16wide with 16/8 wave slots per simd but each simd can access other simd data. This would be similar to rdna3 dual issue but if they could access the data then it would work better in gaming vs rdna3 (+1-6% dual issue gain). Would be similar to GCN but with 6x lower ocupancy levels so no wasted time/latency on waiting. This would simplify the gpu and allow for that 3ghz+?

rtx4000 toping out at 2600mhz on N4 they are packing 32wide. Maybe going 16wide lowers the voltage leak ?

The rtx 40 GPUs are actually able to boost to 2.8ghz just fine, so 3ghz+ for new designs from AMD/Nvidia is not too much of an issue. I still don't see anything above 3.2ghz at similar power consumption for either vendor though. Not without sacrificing something, like SM/CU per clock throughput (like Maxwell vs Kepler).

 

[marees]

Something really must be inbound. Prices dropping in anticipation and to clear existing inventory: https://videocardz.com/newz/amd-radeon-rx-7700-xt-price-has-just-hit-379

I don't think anything is inbound (before September at the earliest)

Its probably a combination of 2 things

1. Navi 22 & Navi 21 stocks getting depleted
2. Navi 32 sales started to decrease (maybe everyone who wants one has already got one 🤔)

 

[branch_suggestion]

Laptops, RDNA3 was effectively culled from the largest volume market which forced RDNA2 to stay in production for way longer than anticipated.
That is probably the most convincing data for RDNA3 being a dog in power sensitive applications.
RDNA4 needs to be pushed as hard as humanly possible in laptops, and the desktop cards need to sell in volume.
For the price, remember AMD will be looking at the gaming segment GM's, and console sales have jumped off a cliff, but that does bring BU GM's up which means that they can charge less for client GPUs, maintain GM's and make up for lost console revenue whilst addressing a weak market position.
Everything is shaping up for a big value generation, they just need to not fumble. Getting good market presence is essential for the follow up to make a splash with the giga parts.
AMD got the first step right with HD 4000, great value, enough performance and they made a lot of them.
For 5000, they had to go for the ~500mm^2 big boy die to take the perf crown and use that strong market foundation for inflating margins. But they failed to go for the kill when they should've known NV would fight back with anything they could muster.
A shame, hopefully they really have got it right this time.

 

[branch_suggestion]

Laptops, RDNA3 was effectively culled from the largest volume market which forced RDNA2 to stay in production for way longer than anticipated.
That is probably the most convincing data for RDNA3 being a dog in power sensitive applications.
RDNA4 needs to be pushed as hard as humanly possible in laptops, and the desktop cards need to sell in volume.
For the price, remember AMD will be looking at the gaming segment GM's, and console sales have jumped off a cliff, but that does bring BU GM's up which means that they can charge less for client GPUs, maintain GM's and make up for lost console revenue whilst addressing a weak market position.
Everything is shaping up for a big value generation, they just need to not fumble. Getting good market presence is essential for the follow up to make a splash with the giga parts.
AMD got the first step right with HD 4000, great value, enough performance and they made a lot of them.
For 5000, they had to go for the ~500mm^2 big boy die to take the perf crown and use that strong market foundation for inflating margins. But they failed to go for the kill when they should've known NV would fight back with anything they could muster.
A shame, hopefully they really have got it right this time.

Oh, and I think it is crucial that they use xx70 as the flagship naming, as full N48 BoM should be closer to the 7700XT than the 7800XT. As should the price.

 

[Mahboi]

I am a bit confused here, as it seems you may be saying contradicting things, at least to me. You say that nothing will replace RDNA 3 at the high end, but then say that an 240mm die will be the high end.

That's why I put high in quotes. It is nothing high end, it's a die the size of a 6600 xt.
It's a card that will compete somewhere between a future 5060 Ti and 5070.
Absolutely unambitious.

Assuming this new die is RDNA 4, either it is faster and it replaces the 7900XTX, or it is slower and it doesn't. It may be newer, but if it isn't faster, then it isn't the new top end card.

It's expected to be above a 7900 xt in raster and around a 4070 Ti in raytracing, making it slightly slower at raster, better at RT, and consuming 215W instead of 350W for the XTX.
I think there won't be many people buying XTXes for gaming once it's out. Nobody really. But XTX could still serve as a compute card for some, however how many people is this?

I suppose I should clarify that by high end, I meant 7900XT and above, mainly thinking the XTX. Still, if the 7900XTX remains the fastest AMD card for some time, then it isn't replaced, so long as there is stock, new or used. I suppose they could stop producing them, and supply could start to dry up, but that doesn't make the 7900XTX any slower. Replacing it would seem a very bad move, if there isn't anything faster.

It's really too close to care.
The real problem is that it's DOA vs Blackwell. It competes with lowest high end cards of Lovelace, 2 years later.

 

[Aapje]

I don't understand why they would keep the 7900 XTX for the current price of $900 and then sell a 8800 XT that is less than 15% slower for let's say $550. Who would ever buy the 7900 XTX for almost double the price, especially when Nvidia will probably just bury the 7900 XTX with a faster 5070 Ti for $800? Or a way faster 5080 for $1k.

Of course they could almost certainly still make a profit by selling the 7900 XTX for $700, but that would still risk them being undercut if Nvidia decides to give a strong price/perf increase. Even in the best case, the average margins would be quite low and they would get into issues selling off the less good dies that they now sell as 7900 XTs and 7900 GREs. They would probably need to keep selling the 7900 XT as well, even though it would be older tech and a little slower than the 8800 XT.

It would just be a mess.

 

[Mahboi]

They were pricing their cards 40+% lower @ similar performance, or pricing it similar but with 40%+ more performance, then?

Ok sorry but can you stop bolding or ALLCAPSING every sentence, it feels like you're putting the finger to what you want me to read when I'm already reading it.

And I don't know where you got the 40% idea. Thermi was a bad era for Nvidia since they risked a lot to get into Compute and won. While AMD tried to run Nvidia out of the market by making an excellent card that they sold at a low price. They ate their own margins to run Nvidia out while Nvidia was in a bad spot, it did nothing but let Nvidia open the Compute market, make their margins almost same as normal, and come back immensely stronger with the fixed Fermi later down the year. While AMD was left having flooded the market with good cards, having made very little margins off of it, and having less resources than Nvidia for the next gen.

If they tried this strategy again now, when Lovelace is absolutely solid, they would get strictly nothing. Just lose money. Reminder that if you sell 1M products with a 100$ margin on each, or you sell 1.2M products with a 84$ margin on each, you have literally made no more money. You just shipped more products.
If AMD had strong market rep, it would be a different story, but right now, lowering prices means nothing. Those who want to buy Nvidia will buy Nvidia anyway. Nvidia doesn't have to lower its prices to compete with AMD since it has legions of convinced shills that will just Defend Geforce's Purity every single day on Reddit. No matter what AMD's prices are, AMD cards are too bad to buy to these brainlets. If the prices are high, AMD is not competitive. If the prices are low, it's because AMD's cards are not competitive and have to be priced at flea market prices. You can't win in a price war with clients that are auto convinced that the only good option is Nvidia. So they just sell high while I sit here simmering in my distaste for this market.

 

[KompuKare]

They undercut Intel MASSIVELY: either vast more performance @ similar price or vastly lower price for similar performance: THAT'S what they need to do in the GPU scene.

Difference is, even a large CPU is generally 200mm² or less, does not ship with VRAM and a board etc. so there simply is far more margings in CPUs to compete.

While it is true that the various "how much does X cost to manufacture" threads generally do not have all the information, irrespective of the actual wafer costs, GPUs simply bigger.

For AMD to come in and undercut Nvidia like Zen 1 did to Intel's HEDT is simply a lot harder. Chiplets are about the only hope but with RDNA3 that didn't work too well.

 

[Mahboi]

AMD is doing the best they can with what they have. Going against established (and manipulative) companies like Intel/Nvidia is not something they can afford to do with an outright price war because both of those companies can fire those shots wayyyy better until AMD has no choice but to bow out of the market. It's not fair to the little guy buying crap from Intel (their space heaters) and Nvidia (proudly and shamelessly selling an overpriced product with reliability issues) but then the little guy refuses to do his research and just subscribes to the herd mentality without question. I don't see things changing any time soon.

AMD can't please anyone in their situation.

If they sell faster stuff at lower prices, they won't be able to handle the volume and people will still buy Intel/Nvidia because those two will come up with clever/deceptive marketing tactics to sell their crap.

If they sell their faster stuff at Intel/Nvidia prices, well, then the consumers won't be happy about that and most of them may still go with Intel/Nvidia.

I just don't see any win-win situation for AMD, other than the enterprise sector where they have built up great loyalty through their power efficiency because every megawatt there counts.

I see one.
AMD needs to make themselves a market identity that's not just copying NVidia. No more pretense of "great RT performance" or such things. Insist on why you put so much VRAM, talk with developers and record it. Papermaster is already doing that with Naffziger, Norrod and even TSMC's Y. J. Mii, there is no way that they can't call Bethesda or someone and go "why did you pick FSR, why do you like open source stuff, did you know that AMD does a lot of open source stuff".

If I could sell myself as a consultant to AMD, literally 90% of my first months would be to try and get the people inside to talk to me and explain how they envision the needs of the market/gamers/developers/enterprise compute and how everyone wants to make tomorrow's gaming happen. Nvidia entirely controls the narrative and AMD is always going to dance one beat behind, so they need to get their own image and vision out.

Then when you have a decent idea of what Radeon does and why, you do as the marketing does, and present it in a good light to everyone.
"this is why the 7800 xt is a great product"
"this is why developers want this much VRAM"
"this is why FSR is made the way it is"
"this is what we want to do for gaming next"
"this is what developers need the most for your games"

And you explain it in a way that shows why AMD is doing it right, even when the product is as poor as RDNA 3.
If Papermaster can take the time to do conversations/interviews with his fellows, surely having a guy with a camera, an editing monkey, and an interviewer (probably 3 times the same guy really) can build a narrative with videos, tweets, shorts etc.
It's not "print all the bus commercials in the city with giant RDNA 5 posters". It's relatively cheap and if functional, a ton of users will finally have arguments to respond to NV's bs narrative. Frankly even just their VRAM gimping is a giant red target on their heads.

 

[Mahboi]

Something really must be inbound. Prices dropping in anticipation and to clear existing inventory: https://videocardz.com/newz/amd-radeon-rx-7700-xt-price-has-just-hit-379

They likely want to remove any unnecessary RDNA 3 out of the way so RDNA 4 can come out with a bang, unlike N33 back in early 23.

I don't think anything is inbound (before September at the earliest)

Its probably a combination of 2 things

1. Navi 22 & Navi 21 stocks getting depleted
2. Navi 32 sales started to decrease (maybe everyone who wants one has already got one 🤔)

I fail to see how that's not inbound?
September isn't that far away, 4 months to clear the remaining RDNA 3 inventory isn't a luxury. If anything, RDNA 2 is still around. (jesus christ)
Also between the 18 Gbps leak on the GDDR6 VRAM, it is extremely likely that AMD does want to go for a large market share here. They are basically going for all the options to serve the highest volume.
But it's AMD so I'll take a rain check on that until the official announcement...

 

[H T C]

If they tried this strategy again now, when Lovelace is absolutely solid, they would get strictly nothing. Just lose money. Reminder that if you sell 1M products with a 100$ margin on each, or you sell 1.2M products with a 84$ margin on each, you have literally made no more money. You just shipped more products.

And took away market share from the competitor in the process, which is what they need the most right now.

Once they have significantly more market share, THEN they can raise prices again.

Difference is, even a large CPU is generally 200mm² or less, does not ship with VRAM and a board etc. so there simply is far more margings in CPUs to compete.

Makes total sense.

Still, and as things currently stand, unless AMD comes up with "a killer GPU", i don't see AMD making headway against nVidia UNLESS they resort to "drastic steps", like what i suggested.

 

[Aapje]

@Mahboi

Indeed. If you only ever copy, you will never be better at anything.

However, asking the existing userbase is never going to result in something truly revolutionary. As Henry Ford perhaps said: “If I had asked people what they wanted, they would have said faster horses.”

Vision needs to come from within.

 

[branch_suggestion]

The real problem is that it's DOA vs Blackwell. It competes with lowest high end cards of Lovelace, 2 years later.

Ehh, that is TBD at least in raw PPA terms.

And I don't know where you got the 40% idea. Thermi was a bad era for Nvidia since they risked a lot to get into Compute and won. While AMD tried to run Nvidia out of the market by making an excellent card that they sold at a low price. They ate their own margins to run Nvidia out while Nvidia was in a bad spot, it did nothing but let Nvidia open the Compute market, make their margins almost same as normal, and come back immensely stronger with the fixed Fermi later down the year. While AMD was left having flooded the market with good cards, having made very little margins off of it, and having less resources than Nvidia for the next gen.

Thermi was just bad engineering, same with bumpgate. AMD made the same push with GCN but lacked the spare resources to make a compute stack, relying on OpenCL and the like to get adopted instead.

If they tried this strategy again now, when Lovelace is absolutely solid, they would get strictly nothing. Just lose money. Reminder that if you sell 1M products with a 100$ margin on each, or you sell 1.2M products with a 84$ margin on each, you have literally made no more money. You just shipped more products.
If AMD had strong market rep, it would be a different story, but right now, lowering prices means nothing. Those who want to buy Nvidia will buy Nvidia anyway. Nvidia doesn't have to lower its prices to compete with AMD since it has legions of convinced shills that will just Defend Geforce's Purity every single day on Reddit. No matter what AMD's prices are, AMD cards are too bad to buy to these brainlets. If the prices are high, AMD is not competitive. If the prices are low, it's because AMD's cards are not competitive and have to be priced at flea market prices. You can't win in a price war with clients that are auto convinced that the only good option is Nvidia. So they just sell high while I sit here simmering in my distaste for this market.

Maxwell was the end of rationality, yeah.
AMD has tried the small>big strat many times, 7000 went alright but Hawaii was really needed, and it was good, but had a terribad cooler that took way too long to replace and was slightly too small.
GCN tried to do compute and gaming but AMD really should've picked gaming only in hindsight, jumping into compute when there was money to be made and they had money to spare.
NV saw Kepler's overblown compute as a weakness that AMD could exploit with a trimmed down gaming gen, and so made Maxwell, which was also made to work at low power in Tegra. AMD thus tried again with Fiji, using HBM as their hail mary but this was once again a compute focused card with no real market yet.
They went back to small big with Polaris, which held market share but was worse economically than Pascal, and then Vega which once again did nothing to change the high end.
NV had big CoWoS/HBM issues with P100 but were able to learn, and now they really felt they were ready for a compute first product after two gens of gaming first.
Volta ruined AMD's only bastion remaining of compute perf, thus they would too have to invest in two types of GPU to compete across the spectrum, but took the bifurcation far further than NV did.
This means far less software compatibility between compute and gaming cards, something NV kept as a high priority to keep devs invested in their ecosystem. But AMD knew that problem would need to be dealt with after they get the raw hardware lead back.

 

[Mahboi]

Ehh, that is TBD at least in raw PPA terms.

Where does the PPA have any connection with the market conditions here?
Yes in transistor/area/production cost vs price, it matters. But the market doesn't care about the prod costs, just the prices. Remember the 2020-2021 era?

Thermi was just bad engineering, same with bumpgate. AMD made the same push with GCN but lacked the spare resources to make a compute stack, relying on OpenCL and the like to get adopted instead.

I see. BTW what happened with OpenCL? They held on to it for so long supposedly, but I've heard that even Nvidia covered OpenCL decently better (I.E OpenCL 3.0 while AMD was stuck with 2.0, etc). What went on there?

much history

Thanks professor, it's nice to learn.

 

[Kepler_L2]

It is so weird. AMD had navi 36 & also a 3d v cache of N31 & also RDNA 3.5 APUs, but no sign of replacement for the 'flawed' N31 parts. I don't get it at all. Hopefully AMD will tease a high level roadmap in June at computex

Cost of respins, new dies etc. is in the many many millions. They don't sell enough volume of high-end GPUs to justify it.

 

[adroc_thurston]

Pricing their cards SLIGHTLY cheaper than nVidia @ same or slightly better performance IS NOT ENOUGH: AMD needs to "aim for the jugular" and either have a similar price for SUBSTANTIALLY HIGHER PERFORMING cards (i mean like 40%+ higher), or MUCH CHEAPER price for similar performance cards (i mean like 40%+ cheaper), EVEN IF they take a monetary loss in the immediate future BECAUSE of it.

"please nuke your margins"
No. Lmao.
It's not a charity.

 

[Aapje]

Yes in transistor/area/production cost vs price, it matters. But the market doesn't care about the prod costs, just the prices. Remember the 2020-2021 era?

The seller side of the market definitely cares about the production costs.

Arguing that the seller side doesn't matter by pointing to a period where demand greatly outstripped supply doesn't prove anything. I can just as easily point to markets with very low demand where sellers exited the market because their costs were higher than their incomes.

 

[branch_suggestion]

The seller side of the market definitely cares about the production costs.

Arguing that the seller side doesn't matter by pointing to a period where demand greatly outstripped supply doesn't prove anything. I can just as easily point to markets with very low demand where sellers excited the market because their costs were higher than their incomes.

Yeah a product is a bust when it cannot maintain volumes at a given margin.
GPUs are like the single most vital product in this aspect.

 

[Mopetar]

I know this is the RDNA4 thread, but from that tweet Navi 50 basically looks like Navi 40. Any idea as to any meaningful differences to expect, because the listed N5X parts just look like big RDNA4 parts or that the current intent is for RDNA4 to stick around and be the the low end parts for the next generation when AMD makes bigger cards.

 

[Saylick]

"please nuke your margins"
No. Lmao.
It's not a charity.

"Please nuke your margins so that Nvidia is pressured to lower their prices, so that I can finally buy that high-end Nvidia GPU I've always wanted but could not afford"

FTFY.

 

[marees]

"Please nuke your margins so that Nvidia is pressured to lower their prices, so that I can finally buy that high-end Nvidia GPU I've always wanted but could not afford"

FTFY.

Intel will do this to AMD with battlemage