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

Win2012R2 · Thursday at 5:52 PM

Jan Olšan said:
AMD said they didn't give that figure

True, but the fact that they say so likely means it's lower - or very little higher, why would not they go on record and say they sold (for example) 1 mln cards in record times ever? All they say it's 10x better than other launches, which isn't something you'd say if you sold 1 mln cards.

Bottom line is this - they obviously did not have enough stock and MSRP is fake too, and maybe that's the way they like it, aha

blackangus · Thursday at 6:17 PM

Win2012R2 said:
Bottom line is this - they obviously did not have enough stock and MSRP is fake too, and maybe that's the way they like it, aha

Its not the bottom line - its an incorrect conclusion.

There were many cards available at MSRP...saying anything else is just intentional mis-representation.
Of course there were many cards not at MSRP which happens every generation as AIB's always have "enhanced" offerings.

The only reason their stock was not enough is because NVidia was MIA nearly completely.
You keep making the statement above while completely ignoring facts and market dynamics, which invalidate the claim.

Why do you feel the need to keep repeating this every few pages? It just comes across like you have some axe to grind.

IMHO
The bottom line is NVidia completely missed on their execution to an un-precidented extent which lead to completely unexpected record AMD sales.

jpiniero · Thursday at 7:36 PM

blackangus said:
There were many cards available at MSRP...saying anything else is just intentional mis-representation.

Only because AMD gave rebates on a few cards. Obviously there isn't much of a point in doing that if they were going to be able to sell the entire (6 months worth?) of inventory instantly.

DaaQ · Thursday at 8:00 PM

Win2012R2 said:
To win market share AMD must make enough cards to satisfy demand - especially for the launch period, but that's clearly has not happened, if 200k units sold is true (seems plausible given known stock figures at a bunch of retailers) then they clearly failed to make anywhere near enough to grab even tiny market share, despite their claims that they want to get to 40-50%, they won't if they got no stuff to sell.

Jan Olšan said:
The 200 000 number was some sort of reporting mistake, AMD said they didn't give that figure. Not sure if BenchLife got it off record (or from another source), but they deleted the article that said that. I would basically forget the number ever made news, probably is way off.

To piggy back off of @Jan Olšan

everyone loves Techpower up for quoting so here.

TechPowerUp

www.techpowerup.com

I guess you could assume that the tenfold comment means they only sold 2000 7900xtx/xt cards at launch. Doubtful, but I can you can assume that. We all know how that saying goes I hope.

While rumors suggested AMD sold 200,000 Radeon RX 9000 GPUs at launch, the company clarified that this figure was never shared, basically dismantling the rumor. When (and if) we get the official concrete sales numbers, it will show just how much of a significant milestone has been achieved by Team Red. Notably, in some specific markets like Japan, AMD has captured nearly 50% of the market share with the RX 9070 series—a first for the company. Dr. Su confirmed that AMD is ramping up production to meet surging demand. "We are very excited and increasing manufacturing to ensure more gamers can access our GPUs," she said.

Google Search

this is just the google search, seem alot is focused on AI product.

but here a quick and dirty context description of the claim.

Nvidia says it's sold twice as many RTX 50-series cards as RTX 40-series in the first 5 weeks. I'd bloody well hope so given there was essentially just the RTX 4090 for competition

So, combined it sold more of four different GPUs than a single $1,600 card.

www.pcgamer.com

The RTX 40-series, on the other hand, only had two separate GPUs launching in those first five weeks. The RTX 4090 launched on October 12 way back in 2022, when my beard was less grey and belly less protrusive, and then five weeks later the RTX 4080 was released.

I know pcgamer but the above link shows the first simple search. I guess if you count "Blackwell" or as "50 series" that makes a difference. The difference is the AI market obviously. Then again I commented on the ridiculous graph they used previously.

moonbogg · Friday at 12:19 AM

Just stopping by to deliver many negative sentiments. Lack of supply this, Nvidia ruined gaming that, just the usual.

del42sa · Friday at 2:08 AM

coercitiv said:
AMD discovered the secret of success in the GPU market: they just need Nvidia to stop making cards.

View attachment 120703

one can't help thinking, that the success of RDNA4 could be even greater if AMD produced a third version of the RDNA4 chip family lets say 450 - 470mm2, which would be still less than VEGA 64 size (510mm2), but AMD never misses an opportunity to miss an opportunity, right ?

ToTTenTranz · Friday at 4:20 AM

del42sa said:
one can't help thinking, that the success of RDNA4 could be even greater if AMD produced a third version of the RDNA4 chip family lets say 450 - 470mm2, which would be still less than VEGA 64 size (510mm2), but AMD never misses an opportunity to miss an opportunity, right ?

Bigger chips = less chips per wafer = lower availability.
People are actually pretty happy with the 9070XT's performance.

What's killing off consumer GPU availability is mostly just wafer supply. It's UMC and Globalfoundries dropping out when things transitioned to EUV, followed by Samsung and Intel failing to compete in the latest nodes' performance and yields.

TSMC is asking whatever the hell they want for high end nodes, which then leaves lower margin products with lower wafer allocation, which then causes an immense scarcity downstream.

Win2012R2 · Friday at 6:29 AM

ToTTenTranz said:
Bigger chips = less chips per wafer = lower availability.

Better margin per chip = order more wafers

That's how it should work

Win2012R2 · Friday at 6:32 AM

blackangus said:
There were many cards available at MSRP...saying anything else is just intentional mis-representation.
Of course there were many cards not at MSRP

Ok, so which one was bigger in terms of availability - MSRP or non-MSRP?

Both make up 100%, so what do you reckon the split is? I'd say 20-80 at best, and MSRP heavily skewed towards less desireable non-XT model.

blackangus said:
The only reason their stock was not enough

I am glad that we both agree on that

GaiaHunter · Friday at 7:23 AM

Win2012R2 said:
Better margin per chip = order more wafers

That's how it should work

Fabs at full capacity = no more wafers to order.

Billion dollars fabs take time to be built.

And that is assuming ASML can produce enough tools for TSMC equip more fabs.

Win2012R2 · Friday at 8:01 AM

GaiaHunter said:
Fabs at full capacity = no more wafers to order.

Pre-Order early?

It's not like it takes 3 months to make a new chip.

The answer to AMDs reluctance is obvious of course - much higher profit margins for using 4nm to make EPYC chiplets, GPUs is just big risk to be humiliated yet again by Nvidia, lose money on depreciating non-competitive overstock, lose opportunity to make extra money via EPYC, so it's no brainer from their point of view - only use capacity that they can't sell for higher margin. It makes sense on greed level, but strategically terrible choice to continue making.

blackangus · Friday at 8:04 AM

Win2012R2 said:
Ok, so which one was bigger in terms of availability - MSRP or non-MSRP?

Both make up 100%, so what do you reckon the split is? I'd say 20-80 at best, and MSRP heavily skewed towards less desireable non-XT model.

No goal switching, now, thats not what you said!

Win2012R2 said:
I am glad that we both agree on that

I was using your words, and you seem to be purposeful missing the point.
In no SANE world would AMD take a risk that would mean they could supply both their normal market and NVIDIA's whole market.

Win2012R2 · Friday at 8:07 AM

blackangus said:
In no SANE world would AMD take a risk that would mean they could supply both their normal market and NVIDIA's whole market.

AMD is at 10% market share, I never said they need to supply 10 times more - that's obviously too risky, but if you are claiming, like AMD does, that they specifically crafted RDNA4 to go for 40-50% market, because that AMD guy is "80% kind of guy" (his words I believe), then making 200k cards (even if it's not exact number this seems like good estimate based on other verified numbers) isn't even good enough for 10% share they have at the moment. Additionally they had at least 2 extra months of production for the launch because the cards were in shops already in January.

So no, they did not make enough stock, for understandable reason of being able to sell same silicon for a lot more money elsewhere, which is why if they said so but their claims of aiming for 40-50% is clearly untrue.

blackangus said:
No goal switching, now, thats not what you said!

If 20% of cards were MSRP then it's fake MSRP, does not matter is such number is "many" - say 40k cards out of 200k, that still means fake MSRP. Now if MSRP was clearly above 50% then I'd say - ok, this is fair - some of the more expensive cards got better coolers etc, fair enough, but no way it was 50% MSRP.

blackangus · Friday at 8:20 AM

Win2012R2 said:
It's not like it takes 3 months to make a new chip.

It is like it take 3 months (or more) to make a new CARD.
Let me break this down a bit (which will still be high level)

AMD: Hey we need to get high volume production.
AMD: Has to get approval to even thing more on this - OK! - Let get this moving.
AMD: How much should we build? (Research/Debate/upper management review, budget planning - 2-3 Weeks right here.)
AMD: Talks to TSMC for chips
AMD: Talks with AIBs to give them lead time to order components
AIB's: Oh no! We were not prepared for this
AIB's: How much should we build? (Research/Debate/upper management review, budget planning - 2-3 Weeks right here.)
AIB's: Start esclated discussions with their supply chain to see lead time on parts orders
AMD: Orders wafers waits on TSMC for wafer arrival
AMD: Possibly waits on TSMC for production slots (Not all wafers will be able to be in production at the same time)
AIB's: Start Parts orders and wait
AMD: Start validating initial chips
AMD: Start shipping initial chips to AIB's
AIB's: Waiting on parts but have a small chip supply
AIB's: Production started and completed
AIB's: Ship to distrbuters
Distys: Ship to Vendors:
Small amount of cards available over original intent.

When I order enterprise hardware it takes 2-3 months to get something already built and not something specialized.
Some hardware we order that is specialized take 6-12 months for delivery because its in demand. And the company we get that from is MUCH larger than AMD.

blackangus · Friday at 8:25 AM

Win2012R2 said:
AMD is at 10% market share, I never said they need to supply 10 times more - that's obviously too risky, but if you are claiming, like AMD does, that they specifically crafted RDNA4 to go for 40-50% market, because that AMD guy is "80% kind of guy" (his words I believe), then making 200k cards (even if it's not exact number this seems like good estimate based on other verified numbers) isn't even good enough for 10% share they have at the moment. Additionally they had at least 2 extra months of production for the launch because the cards were in shops already in January.

So no, they did not make enough stock, for understandable reason of being able to sell same silicon for a lot more money elsewhere, which is why if they said so but their claims of aiming for 40-50% is clearly untrue.

AMD never said they were going to change the market in 1 quarter, that is a false assumption which your are ignoring. Getting that much market share takes time and they know it. So the above doesn't in any way justify the insane finacial risk that comes with it. No real business would EVER take that risk. (Start up maybe)

Win2012R2 said:
If 20% of cards were MSRP then it's fake MSRP, does not matter is such number is "many" - say 40k cards out of 200k, that still means fake MSRP. Now if MSRP was clearly above 50% then I'd say - ok, this is fair - some of the more expensive cards got better coolers etc, fair enough, but no way it was 50% MSRP.

This is goal moving. Cards were available at MSRP, and the AIB's enhanced ones were also available at their MSRP. This is no different than any previous generations from either vendor.

Win2012R2 · Friday at 8:34 AM

blackangus said:
Cards were available at MSRP, and the AIB's enhanced ones were also available at their MSRP

A small amount of cards (likely not even 10%) were at MSRP as declared by AMD, which makes MSRP as declared by AMD - fake, which is what I said.

Just because AIBs had inflated AMD's MSRP and sold it at exactly their own MSRP does not mean that AMD's MSRP was not fake - it was.

blackangus said:
AMD never said they were going to change the market in 1 quarter, that is a false assumption which your are ignoring

I am focusing on what they said, and the fact that they failed to mention that 40-50% aspiration is by (for example) 2030 is on them deceiving with their statements.

RDNA4 is a very good product and it's here for at least 18 months, possibly even longer since 4N is going to be a cheaper than 3-2N and have more capacity.

Therefore IF AMD actually had real intention to work towards their declared market share goal then they would have made a LOT of cards for launch - like at least 1 million, which incidentally is just around 10% dGPUs sold PER QUARTER. They'd have to make 2-3 mln for launch if they were really serious, and sell it at real MSRP - and that stuff (9070 XT) would have sold!

branch_suggestion · Friday at 8:51 AM

Win2012R2 said:
I am focusing on what they said, and the fact that they failed to mention that 40-50% aspiration is by (for example) 2030 is on them deceiving with their statements.

Assume any long term AMD plan is 5 years, so in this case 2030 is spot on.
Or really 3 product generations, if GFX13 is a banger then through GFX14 AMD could really be around the 40% mark.
Especially if NV decides to cheap out on client with 18A-P for lower end BW-Next.
NV isn't crazy, high end BW-Next should be TSMC surely.

Kronos1996 · Friday at 11:01 AM

To actually achieve that marketshare they need to be in OEM Laptop and Desktop as well. Seems like RDNA 3 killed that momentum and RDNA 4 doesn’t even have design wins despite being good. Are laptop makers just completely done or are they more interested in Halo APU’s? Will those even count as GPU marketshare?

Jan Olšan · Friday at 11:48 AM

del42sa said:
one can't help thinking, that the success of RDNA4 could be even greater if AMD produced a third version of the RDNA4 chip family lets say 450 - 470mm2, which would be still less than VEGA 64 size (510mm2), but AMD never misses an opportunity to miss an opportunity, right ?

A theory I have is that if AMD cancelled both their highend and mainstream version of the GPUs (only Navi 44 is left from the original lineup) in early 2023, they possibly didn't have time for highend version any more.

Gotta remember that Navi48 is also a new design, not an originally intended GPU. Perhaps they had a look at their resources (team headcount, the total manhours available till the shipping target) and it was clear they can only do a single new monolithic RDNA 4 design in time, to accompany the Navi 44 that was already being worked on. Perhaps it was simply a choice - make Navi 48 and cover the lower parts of the market where people are more cost sensitive and the market share opportunity is likely much better - or do a highend while having nothing in the mainstream? Highend buyers are more likely to ignore AMD despite much better bang-for-buck ratio and always go Nvidia no matter what.

If you only could make one of the two, I think Navi 48 was the prudent choice over Navi 4highend.

GodisanAtheist · Friday at 12:28 PM

Jan Olšan said:
A theory I have is that if AMD cancelled both their highend and mainstream version of the GPUs (only Navi 44 is left from the original lineup) in early 2023, they possibly didn't have time for highend version any more.

Gotta remember that Navi48 is also a new design, not an originally intended GPU. Perhaps they had a look at their resources (team headcount, the total manhours available till the shipping target) and it was clear they can only do a single new monolithic RDNA 4 design in time, to accompany the Navi 44 that was already being worked on. Perhaps it was simply a choice - make Navi 48 and cover the lower parts of the market where people are more cost sensitive and the market share opportunity is likely much better - or do a highend while having nothing in the mainstream? Highend buyers are more likely to ignore AMD despite much better bang-for-buck ratio and always go Nvidia no matter what.

If you only could make one of the two, I think Navi 48 was the prudent choice over Navi 4highend.

- Yeah, N48 suggests it was the last die to be designed in the stack, so a sort of backfill die after N41/42/43 got scraped. Makes sense that AMD basically built the thing as lean and mean as they could get away with because it would be doing a lot of heavy lifting across several product categories this gen.

del42sa · Friday at 12:31 PM

Jan Olšan said:
A theory I have is that if AMD cancelled both their highend and mainstream version of the GPUs (only Navi 44 is left from the original lineup) in early 2023, they possibly didn't have time for highend version any more.

Gotta remember that Navi48 is also a new design, not an originally intended GPU. Perhaps they had a look at their resources (team headcount, the total manhours available till the shipping target) and it was clear they can only do a single new monolithic RDNA 4 design in time, to accompany the Navi 44 that was already being worked on. Perhaps it was simply a choice - make Navi 48 and cover the lower parts of the market where people are more cost sensitive and the market share opportunity is likely much better - or do a highend while having nothing in the mainstream? Highend buyers are more likely to ignore AMD despite much better bang-for-buck ratio and always go Nvidia no matter what.

If you only could make one of the two, I think Navi 48 was the prudent choice over Navi 4highend.

I agree, but if you want to regain lost territory (meaning market share) you have to do anything except making a prudent choice

jpiniero · Friday at 12:52 PM

blackangus said:
When I order enterprise hardware it takes 2-3 months to get something already built and not something specialized.
Some hardware we order that is specialized take 6-12 months for delivery because its in demand. And the company we get that from is MUCH larger than AMD.

Like just getting the finished product from China to the US can take a month, if shipped by boat.

The 5070 is showing actual availability now so we'll have to see if that causes the 9070 Non XT sales to slump.

adroc_thurston · Friday at 12:58 PM

GodisanAtheist said:
after N41/42/43 got scraped.

40/41/42. 43 never really existed.

blckgrffn · Friday at 2:07 PM

jpiniero said:
Like just getting the finished product from China to the US can take a month, if shipped by boat.

The 5070 is showing actual availability now so we'll have to see if that causes the 9070 Non XT sales to slump.

Yeah, I don't know. Unless its the $550 5070s, the hate I see on Slickdeals and Reddit is strong for a card that at best barely beats a 4070S. The disdain for the 12GB frame buffer (even if "fine" today) is super strong and I can't wait until nvidia is awkwardly positioning a card that might be stronger than a 3070 Ti (the 5060 Ti) with more vram at a slightly lower price,

Interesting times, AMD might have stumbled into a bit of a winner here. Which I think everyone can agree is probably healthy for the long term viability of both companies competing in this space.

DAPUNISHER · Friday at 3:19 PM

Make no mistake, AMD has a winner in RDNA 4.

Nvidia shills are losing the propaganda war this round. On Reddit I see them getting sent to downvote hell for pushing bad talking points. The bad drivers talking point that was the bread and butter for so long backfires spectacularly now.

There is no image management or counter argument that is going to change gamers minds about vram. Green brought nothing in the way of bang for buck this round, and cheesed on vram again. No amount of loserbenchmark rhetoric is going to convince gamers otherwise.

The worst shill tactic I've read there, is fake Radeon owners claiming if their RDNA 3 card does not get FSR 4 they are never buying AMD again. 🤣 Nvidia leaves owners behind on features almost every gen. But if AMD does it just once, they are switching to the company that does it habitually. GTFO with your faux outrage. Remember how they roasted AMD for making FSR non proprietary? Pepperidge Farm remembers.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Senior member

Lifer

Golden Member

Lifer

Member

Senior member

Senior member

Senior member

Diamond Member

Senior member

Senior member

Senior member

Senior member

Senior member

Senior member

Senior member

Member

Senior member

Diamond Member

Member

Lifer

Diamond Member

Diamond Member

Super Moderator CPU Forum Mod and Elite Member