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

 

[Ajay]

PC is a bad market to be in.

Yeah, aside from Apple, it’s a low margin business. HP and Dell rely on service contracts with enterprise businesses partners to generate profits. Consumer has terrible margins only offset by cheap components and very large unit purchases on said components. AMD needs to keep pushing to get into the enterprise segment and rely less on consumer - though that’s not much better, since Intel is giving their silicon away.

 

[jpiniero]

If there's one thing that helps out AMD... it's that you can be pretty sure that the low end Blackwell gaming won't be a big upgrade over Ada. Remember that there's literally no SRAM/IO scaling on N3E compared to N4... and the wafers cost substantially more. The die sizes of GB206 and GB207 almost have to be smaller.

 

[Joe NYC]

PC is a bad market to be in.

Just because AMD mis-played it, does not make it a bad market to be in. Just a year ago, AMD had 2.1 billion revenue from client and it was AMD's #1 segment.

The problem is AMD is waving a white flag, making excuses even before the shooting starts. The competitor then know he can get away with murder. Which is what happened.

 

[Ajay]

Just because AMD mis-played it, does not make it a bad market to be in. Just a year ago, AMD had 2.1 billion revenue from client and it was AMD's #1 segment.

The problem is AMD is waving a white flag, making excuses even before the shooting starts. The competitor then know he can get away with murder. Which is what happened.

Murder? Intel is committing suicide with net negative margins. You want AMD to compete with THAT? Brilliant plan.

 

[jpiniero]

Murder? Intel is committing suicide with net negative margins. You want AMD to compete with THAT? Brilliant plan.

Client margins are probally OK. Desktop ASP was even up compared to last year. 10 nm is dirt cheap.

AMD can't compete because of TSMC's prices.

 

[Ajay]

Client margins are probally OK. Desktop ASP was even up compared to last year. 10 nm is dirt cheap.

10nm? I suppose you mean the new '7nm'. Okay, good to know - so it’s only server CPUs that are net negative?

 

[DisEnchantment]

It would not surprise me if others pay 9.5k per N6 wafer, but AMD pays less.

Much doubt my friend. We are not paying anywhere close to that (we are using 'N7 family' ). And folks using N7A would pay for such things? are you sure? Do people know they are counting cents in that particular industry? You would be well advised to doubt this specific number.

 

[Joe NYC]

The issue AMD had was that even if they halved the price of the chips the cost of the ram and motherboards made switching to AM5 at that time rather expensive. It was the platform cost that was prohibitive and going dirt cheap on pricing wouldn't have changed that by enough.

In addition that kind of play makes AMD look like the value player which means it is always expected that AMD is cheaper even with the better product and AMD don't want to end up in that situation again. Look at Radeon Vs Nvidia, even before stuff like DLSS and Ray Tracing were a thing it was often thought that AMD cards are there to lower NVs price. AMD even went after the small die value play with 4000 and 5000 series, it failed because while they did grab some market share it wasn't enough to make the low margins work and NV responded with a $200 price cut on their GTX 280.

If the 7900XTX launched at $600 for example then NV would cut the cheaper to manufacture 4080 to $800 and everybody would still buy the NV card.

If AMD are going to do deals and offer cheaper parts it will be to OEMs to get into desktop and laptop products, it won't be in the DIY space.

DDR5 prices were the highest when Alder Lake launched.

It did not exactly deter Intel from a bait and switch of sending DDR5 for benchmarking and selling DDR4. Bait and switch.

At the same time, AMD disarmed itself by not doing the same - sending 5800x3d for benchmarking and then selling 5800x + DDR4 in volume in its place, in a similar bait and switch. AMD unilaterally disarmed in client, something Intel did not do.

Intel is so serious about competing that it goes to grotesque lengths, such selling 250 Watt CPUs, just to make the sale, make it to the top of the review charts. AMD would not even paper launch the 5800x3d to prevent Intel / Alder Lake from taking Halo effect from AMD. Until it was too late and eventually, it was launched in the middle of the night, without anyone noticing.

After DDR5 and platform issues are were no longer an issue, AMD still had a weaker rebound in Q2 then Intel. Even from such catastrophic Q1 levels.

In GPUs, I am not faulting AMD efforts to sell a competitively weaker products. It was an unfortunate execution, and AMD can't threaten NVidia with any kind of price war, because NVidia is rock solid financially.

Unlike Intel, which is weak financially, and can be threatened.

 

[adroc_thurston]

Just because AMD mis-played it, does not make it a bad market to be in. Just a year ago, AMD had 2.1 billion revenue from client and it was AMD's #1 segment.

The problem is AMD is waving a white flag, making excuses even before the shooting starts. The competitor then know he can get away with murder. Which is what happened.

That was when PC was hot and they had nice ASPs.
Now you need to be as cheap as Intel Raptor Lake which is something AMD can't quite do.

 

[Joe NYC]

Yeah, aside from Apple, it’s a low margin business. HP and Dell rely on service contracts with enterprise businesses partners to generate profits. Consumer has terrible margins only offset by cheap components and very large unit purchases on said components. AMD needs to keep pushing to get into the enterprise segment and rely less on consumer - though that’s not much better, since Intel is giving their silicon away.

Smart Phone business is even worse, yet, Qualcomm, Media Tech are still relying very heavily on that segment.

AMD has the enterprise to share some costs with, and is still whining about client.

I see you also seem ok with AMD unilaterally surrendering as a response to Intel giving away their silicon cheaply - instead of making this twice as expensive on Intel, forcing Intel to go to the bond market to borrow money to execute this strategy.

AMD just raised the white flag, did not counter, and let Intel get away with this, let Intel keep its fabs from being idle at very low cost.

It also gives Intel confidence that it can go back to the bad old days of strong arming all of the participants in the market place, that AMD will not stand up to it, and it will not stand up for the OEMs selling AMD products.

 

[Joe NYC]

Murder? Intel is committing suicide with net negative margins. You want AMD to compete with THAT? Brilliant plan.

Intel initiated it, so make Intel pay double the cost per unit.

And since Intel sells about 6 units to 1 AMD unit, make that cost Intel 6x in dollar terms.

After that, preventing Intel from taking that market share, let Intel take additional charges for Fab under-utilization.

And in process, of competing, help out the OEMs who were loyal to AMD, instead of abandoning them.

AMD was very well positioned for the price war, but did not show up on the battlefield. Forfeited.

 

[Joe NYC]

Client margins are probally OK. Desktop ASP was even up compared to last year. 10 nm is dirt cheap.

AMD can't compete because of TSMC's prices.

That's not correct any more. Intel fabs are not very efficient, and they are under utilized.

Lisa could also have gone to Taiwan and discussed the situation. That Intel is conspiring to take take away 10s of thousands of TSMC wafer starts by monopolistic market manipulation. And ask TSMC if it wants to do anything about it. Mount a joint effort to stop it from happening.

Stop transferring under-utilization from Intel fabs to TSMC fabs.
(and we know that TSMC is suffering under utilization)

If AMD did not just raise white flag, those costs would go even higher for Intel, because the Intel fabs have ongoing costs, even when half empty. Doubling the cost when half empty.

 

[Joe NYC]

Much doubt my friend. We are not paying anywhere close to that (we are using 'N7 family' ). And folks using N7A would pay for such things? are you sure? Do people know they are counting cents in that particular industry? You would be well advised to doubt this specific number.

The various price estimates I have seen show that that double that price - 15k to 20k - buys you an N5 wafer.

IIRC, from Daniel Nenni, it does not seem that TSMC has a price list and volume discount chart. TSMC negotiates each deal individually on its own, especially large orders for large customers.

 

[Joe NYC]

That was when PC was hot and they had nice ASPs.
Now you need to be as cheap as Intel Raptor Lake which is something AMD can't quite do.

AMD steadily built up the market share, through good times and bad, to be there, to be able to benefit from those nice ASPs.

By abandoning the market and OEMs, AMD will be nowhere near the same position to benefit from the rebound.

That miscalculation of giving Intel 5+ years of hard won market share gains in one fell swoop will be biting AMD in the ass for the next 5 years...

 

[adroc_thurston]

By abandoning the market and OEMs

They didn't.
They undershipped and that was communicated quite clearly.

 

[Joe NYC]

They didn't.
They undershipped and that was communicated quite clearly.

We will see how it plays out.

Undershipping means the same OEMs have Intel inventories burning in their storage, and the OEMs are forced to push Intel (over AMD), just to get rid of that inventory.

And pushing Intel over AMD may then become a habit...

 

[adroc_thurston]

Undershipping means the same OEMs have Intel inventories burning in their storage, and the OEMs are forced to push Intel (over AMD), just to get rid of that inventory

Yea Intel was flooding the channel for a while trying to keep their fabs busy.
IDM things.

 

[SolidQ]

If the 7900XTX launched at $600 for example then NV would cut the cheaper to manufacture 4080 to $800 and everybody would still buy the NV card.

Not everybody, i would go to 7900XTX, because it's faster and cheaper. About everybody would buy NV, look at 3050 vs 6600, people buying 3050, which expensive and slower around 30%, so Huang just can put label "RTX" and people like zombie will buy this, if it's 3x time slower than AMD.

 

[adroc_thurston]

Not everybody, i would go to 7900XTX, because it's faster and cheaper. About everybody would buy NV, look at 3050 vs 6600, people buying 3050, which expensive and slower around 30%, so Huang just can put label "RTX" and people like zombie will buy this, if it's 3x time slower than AMD.

Yes which is why AMD needs an untouchable halo part.
Thank you.

 

[igor_kavinski]

Yes which is why AMD needs an untouchable halo part.
Thank you.

Goes without saying, that it should have zero quality control issues, like the high hotspot temps, insufficient liquid in vapor chamber and Powercolor cooler issues.

 

[adroc_thurston]

Goes without saying, that it should have zero quality control issues, like the high hotspot temps, insufficient liquid in vapor chamber and Powercolor cooler issues.

meh that's board-level whatever.
It just gotta have the longest bars in all reviews and they're gucci

 

[A///]

Gucci make luxury goods and clothing...

ad rock come back to the zen 5 thread and leak some more ocmpany details.

 

[Joe NYC]

Yes which is why AMD needs an untouchable halo part.
Thank you.

Yeah,

And Navi4x turned Navi5x looks to be that.

In the meantime, AMD just has to go with the flow, just maintain enough of the market share to stay in the game, without making any big waves...

Hopefully, AMD will get to at least a worthwhile successor to 7700 / 7800 with RDNA4

 

[TESKATLIPOKA]

Hopefully, AMD will get to at least a worthwhile successor to 7700 / 7800 with RDNA4

What would that entail?
Higher performance?
Higher perf/$?
Higher perf/W?
Or everything mentioned above?
And by how much? 20%, 25%, 33% or more?

200-250mm2 die size isn't that bad.
I would say 2x higher specs compared to N33 should be achievable, but after n31 I am not really sure. It consumed a ridiculous number of transistors for those specs. We will see, I think AMD will want to release next year.

 

[adroc_thurston]

It consumed a ridiculous number of transistors for those specs.

Xtor counts can be whatever you imagine them to be.
Die/total Si area is what matters.