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

 

[adroc_thurston]

I guess that you are just trolling now?

your new is showing.

Things that happened during previous generations typically stay roughly the same.

no.
Most proper gens from AMD before they went broke were 4 to 5 parts.

because there is no Navi 36 chip.

your new is showing^2.

 

[Aapje]

@adroc_thurston

You could also just provide a link to show us what you are talking about, rather than be the way you are being. But I don't think you can, because Navi is relatively recent, so how can there be an old Navi 36 chip?

Most proper gens from AMD before they went broke were 4 to 5 parts.

But I was referring to the generations after that...

 

[adroc_thurston]

so how can there be an old Navi 36 chip?

it's not old lmao.

But I was referring to the generations after that...

And we're BVCK to proper GPU wars. Capiche?

 

[Aapje]

@adroc_thurston

Still no link with evidence.

 

[SolidQ]

Interesting RDNA5 + arm/zen6?

Xbox Moving 'Full Speed Ahead' on Next Gen Console - IGN

Microsoft is moving "full speed ahead" on its next generation console, an internal email from Xbox president Sarah Bond has revealed.

www.ign.com

 

[dr1337]

Interesting RDNA5 + arm/zen6?

Xbox Moving 'Full Speed Ahead' on Next Gen Console - IGN

Microsoft is moving "full speed ahead" on its next generation console, an internal email from Xbox president Sarah Bond has revealed.

www.ign.com

2028 is still pretty far away, idk how they can really call that full speed

 

[Mahboi]

We're not even certain that XBOX will put out a new console.
Last he was seen, Phil Spencer looked like he was speaking under duress. Like management was forcing him to accept giving up on XBOX's future by holding his family hostage or something.

 

[SolidQ]

They already thinking it's ~4070ti super

Top RDNA 4 GPU to overtake RTX 4070 Ti SUPER despite being smaller and using less power per leaked Navi 48 and Navi 44 specifications

Specifications for the RNDA 4-based Navi 44 and Navi 48 have leaked. In the case of Navi 48, AMD might have a killer mid-range chip on its hands as the RDNA 4 GPU is rumored to not only enjoy more raw horsepower than an RTX 4070 Ti SUPER but also use less power and a smaller die size.

www.notebookcheck.net

 

[Mahboi]

I hadn't realised how tiny it is. 240mm². A 6600 xt was 237mm².
We're talking about an entry-midrange chip tied to an oversized 256 bit bus. For comparison the 4070 Ti is 294mm².
Even if it's the most unambitious, tiny thing ever, it is going to be a real kicker when it reaches above 7900 XT performance in raster and XTX/near 4070 Ti in RT.

I'm kind of giving up on my "maybe it'll be bigger RT improvement than that" copium, I don't think a chip this smol is going to reach into 4080 Super RT perf. If it is even just around 4070 Ti, even below it, it'll be impressive.
The last question is how well it'll handle PT/very high throughput RT. Does it just fall apart like RDNA 3, or does it hold up well when heavily queried, thanks to that BVH walker?

Also, considering that the 6600 xt was on 170W and it is on 215W, I'm really expecting some good clocks. Not just 3Ghz. 3.2, 3.3 maybe.

 

[Timorous]

I hadn't realised how tiny it is. 240mm². A 6600 xt was 237mm².
We're talking about an entry-midrange chip tied to an oversized 256 bit bus. For comparison the 4070 Ti is 294mm².
Even if it's the most unambitious, tiny thing ever, it is going to be a real kicker when it reaches above 7900 XT performance in raster and XTX/near 4070 Ti in RT.

I'm kind of giving up on my "maybe it'll be bigger RT improvement than that" copium, I don't think a chip this smol is going to reach into 4080 Super RT perf. If it is even just around 4070 Ti, even below it, it'll be impressive.
The last question is how well it'll handle PT/very high throughput RT. Does it just fall apart like RDNA 3, or does it hold up well when heavily queried, thanks to that BVH walker?

Also, considering that the 6600 xt was on 170W and it is on 215W, I'm really expecting some good clocks. Not just 3Ghz. 3.2, 3.3 maybe.

N48 looks like N10 all over again in terms of relative positioning and maybe price (inflation adjusted).

 

[Mahboi]

N48 looks like N10 all over again in terms of relative positioning and maybe price (inflation adjusted).

"Its price at launch was 399 US Dollars."(TPU)
I'd love it, but there's absolutely no way that AMD will spit on those juicy margins when they'll effectively be selling a 215W 7900 xt+ with an XTX's raytracing. Or a 4070 Ti's rather.
It'll be $600, at least. And for all the madness I don't think inflation has been 50% yet since 2019.

 

[Aapje]

Better a large bus than a small one like Nvidia gives, where the cards choke at higher resolutions.

 

[Timorous]

"Its price at launch was 399 US Dollars."(TPU)
I'd love it, but there's absolutely no way that AMD will spit on those juicy margins when they'll effectively be selling a 215W 7900 xt+ with an XTX's raytracing. Or a 4070 Ti's rather.
It'll be $600, at least. And for all the madness I don't think inflation has been 50% yet since 2019.

The original announced price was $450 or $500 for the anniversary addition. Taking the $450 that would be about $550 now and with it looking slightly more competitive than the 5700XT was then yea I can see $600, I said as much in my guess table.

 

[soresu]

N48 looks like N10 all over again in terms of relative positioning and maybe price (inflation adjusted).

If it's anywhere near the rumoured 7900 XT target for raster gfx then it's a bit above where RX 480 was in positioning, and if the rumoured revamp for BVH units is true it could be way beyond 7900 XTX for ray tracing.

I guess only time will tell, and I'm far less inclined to trust AMD's word on it after the way over optimistic projections in the RDNA3 PR slides.

 

[Timorous]

If it's anywhere near the rumoured 7900 XT target for raster gfx then it's a bit above where RX 480 was in positioning, and if the rumoured revamp for BVH units is true it could be way beyond 7900 XTX for ray tracing.

I guess only time will tell, and I'm far less inclined to trust AMD's word on it after the way over optimistic projections in the RDNA3 PR slides.

Those RDNA 3 PR slides were such a disgrace IMO. AMD had spent years building trust with their marketing decks to just dump it down the drain by massively overpromising with RDNA 3.

 

[Mahboi]

Those RDNA 3 PR slides were such a disgrace IMO. AMD had spent years building trust with their marketing decks to just dump it down the drain by massively overpromising with RDNA 3.

I've been thinking the same for a long time...but as a marketing oriented guy myself, it is a bit tantalizing to think "what would I have done?"
I like to think of myself as very honest, but here were the options:
A) lie (they chose that)
B) tone down to a reasonable positive, talk about "35% improvement" and don't mention power efficiency
C) make an internal deal with management that a respin/RDNA 3.5 will come out within a year and lie
D) make an internal deal and don't lie, and entirely kill RDNA 3 for a whole year because nobody's gonna buy cards that are duds anyway
E) advocate to delay by 6 months to a year for the respin to come out

I like to think E would have come through with sheer will, sweat, and my power to convince Lisa (lol), but really, there is no reason to believe it.

RDNA 3 was crippled. It doesn't matter what you do or say, because any of the options will be painful. Ship a dud, lie, and be a Jensen: marketing will get flak, but the product will ship, you'll make some money, sure it's less than it could've been, but still quite a bit. Ship a dud and don't lie, and everyone will know it's a dud day one. If honesty was rewarded in politics or marketing, people would've started seeing honest politicians and used car salesmen centuries ago. It would have sold nothing and would have hurt the company name far harder, but "marketing would walk away clean". Sorry, but marketing's good name means nothing. The company means everything. It's just how business works, much as I dislike it.

The respin option would also have been painful in several ways. First off, people that bought RDNA 3 would be quite angry at RDNA 3+, because it would have spelled out how much of a dud they bought. Betrayed customers aren't returning customers. (Well maybe I am, but maybe I am a fool. And maybe I just dislike Jensen trying to manipulate me and use FOMO and play with people's inferiority complex that much)

The delay IMO would've been the best course. RDNA 3 would have sold probably as much in the end, if not more. AMD was on the way up with RDNA 1/2, if 3 was late but good, it would've worked out better than just hiding the problems and acting innocent.

I'm not sure of why they didn't delay 6 months. Yes respins are expensive, yes 6 months with nothing but RDNA 2 to sell was a huge stinker. So what. It was a dud anyway.
I fear that the reason they boldly lied was that they felt that it was time to clean the clocks at RTG, as branch_suggestion said. The dud was an alarm bell, it was time to go in hard and wipe out whatever didn't work at RTG. That's the noble assessment.
The non noble one is just that AMD is like Nvidia, they don't care one bit about consumer GPU and just wanted to ignore the problem, move on to CPU and AI, and shipping a dud was not a big concern.

 

[Tuna-Fish]

2028 is still pretty far away, idk how they can really call that full speed

High-end chips generally take more than 4 years to design and manufacture, starting from scratch. If they intend to have a new console with bleeding-edge tech out in 2028, the project had to start last year at the latest.

I hadn't realised how tiny it is. 240mm². A 6600 xt was 237mm². We're talking about an entry-midrange chip tied to an oversized 256 bit bus.

N7 -> N4P.

I think the "oversized bus" makes sense because it uses past-gen GDDR for cost-efficiency reasons, you need the higher width to match the increased compute resources if the memory isn't getting any faster.

Looks really a lot like Polaris to me.

 

[Mahboi]

N7 -> N4P.

I think the "oversized bus" makes sense because it uses past-gen GDDR for cost-efficiency reasons, you need the higher width to match the increased computer resources if the memory isn't getting any faster.

Looks really a lot like Polaris to me.

I have this friend who has been wanting to change his GPU forever.
He's still running an rx 480. Even ran RDR2 at a whopping 18 fps average on it.
Clearly, this GPU was made for him.

 

[Aapje]

The delay IMO would've been the best course.

A delay was not possible since they thought that they could fix it in the drivers, and they were already in volume production. So a respin without releasing the buggy RDNA3 would have meant tossing out all those chips at great cost.

 

[branch_suggestion]

Overall RDNA3 was still mildly successful overall, they made up market share and increased ASP's.
Issue was dGPU volumes this gen were way down on previous, many reasons for that of course.
AMD made the right call, keep moving forward and make do with what you have instead of getting bound up over the present.
I also believe that neutering RDNA4's lineup for TTM and to accelerate RDNA5 are going to pay dividends, dGPU isn't going to see a proper sales bump until something big happens.

 

[ToTTenTranz]

Those RDNA 3 PR slides were such a disgrace IMO. AMD had spent years building trust with their marketing decks to just dump it down the drain by massively overpromising with RDNA 3.

Yes, this particular slide was pretty jarring to be honest, particularly the part where they claim "Architected to exceed 3GHz - Industry 1st" on chips that never really managed to exceed 2.7GHz on gaming workloads.





And that 1.8x RT performance @2.5GHz... Perhaps they were counting on a much larger boost from the double-pumped ALUs? It's hard to say.

The respin option would also have been painful in several ways. First off, people that bought RDNA 3 would be quite angry at RDNA 3+, because it would have spelled out how much of a dud they bought. Betrayed customers aren't returning customers.

I disagree. There are no bad products, only bad prices. And bad names.

They could have released the initial N31 as 7900 and 7900XT, and then the N31+ as 7950XT and 7950XTX, for example. Or they could have released N31+ as 8900XT and 8900XTX (similar to RX480 -> RX580, which didn't seem to bother anyone at the time).

I also believe that neutering RDNA4's lineup for TTM and to accelerate RDNA5 are going to pay dividends, dGPU isn't going to see a proper sales bump until something big happens.

I'm hopeful that RDNA5 is sped up, but it's a bold assumption. Even more if RDNA5 is dependent on a new foundry process that will compete with all the AI demand.

 

[SolidQ]

Even ran RDR2 at a whopping 18 fps average on it.

Say him to try Hub settings

 

[Aapje]

There are no bad products, only bad prices.

That is only true from the consumer's perspective, if you are blind to what companies need to survive (in the long term).

A product whose value is low (relative to production costs, the competition, etc) in the eyes of consumers is a bad product, because the company can't sustainably sell products like that.

At the end of the day, we are all best off if companies make products that have lots of value to us and preferably cost them much less to make than how much we value it.

 

[branch_suggestion]

Yes, this particular slide was pretty jarring to be honest, particularly the part where they claim "Architected to exceed 3GHz - Industry 1st" on chips that never really managed to exceed 2.7GHz on gaming workloads.

It does exceed 3Ghz in games, just need to crank the power way above TDP.

I'm hopeful that RDNA5 is sped up, but it's a bold assumption. Even more if RDNA5 is dependent on a new foundry process that will compete with all the AI demand.

Wafers are not the bottleneck, main bottleneck is HBM now, then adv packaging.
RDNA5 is a lot of things, it's time will come.

 

[SolidQ]

Reddit so funny. RDNA 3.5 is 4070ti super

No, I cannot, because there is no Navi 36 chip. If you evidence that you are not just openly lying here, present it.

N36 was discussed here. Only i don't remember topic. RDNA3 or RDNA4