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]

So FSR 4 with AI upscaling should have back compat to RDNA 3 but may not be effective on RDNA 2 ??

AMD RDNA3 GPUs Can Have A Lot More Vector Registers Than RDNA2 - Phoronix

www.phoronix.com

Also implications for FSR 4 on strix halo ??

Strix Point Variants​

LLVM changes show two Strix Point variants, gfx1150 and gfx1151. ROCm CLR code comments indicate gfx1150 is “Strix,” and gfx1151 is “Strix Halo.” Strix Halo is likely a higher end version of RDNA 3.5.

The most notable difference between the two variants is that gfx1151 (Strix Halo) has the large vector register file found in high end RDNA 3 products. That is, each SIMD has 192 KB of vector registers. gfx1150 (Strix) has a smaller 128 KB vector register file like lower end RDNA 3 parts.

AMD RDNA 3.5’s LLVM Changes

Integrated graphics have been a key part of AMD’s strategy ever since they bought ATI.

chipsandcheese.com

 

[soresu]

Also implications for FSR 4 on strix halo ??

Strix Point Variants​

LLVM changes show two Strix Point variants, gfx1150 and gfx1151. ROCm CLR code comments indicate gfx1150 is “Strix,” and gfx1151 is “Strix Halo.” Strix Halo is likely a higher end version of RDNA 3.5.

The most notable difference between the two variants is that gfx1151 (Strix Halo) has the large vector register file found in high end RDNA 3 products. That is, each SIMD has 192 KB of vector registers. gfx1150 (Strix) has a smaller 128 KB vector register file like lower end RDNA 3 parts.

AMD RDNA 3.5’s LLVM Changes

Integrated graphics have been a key part of AMD’s strategy ever since they bought ATI.

chipsandcheese.com

You seem to be forgetting Kraken Point here.

Differing gfx numbers do not always represent different versions of the µArch.

Usually there are 1-3 numbers representing the µArch 'family', and then another 1-2 to represent the separate die members of the family.

Case in point gfx1200 and gfx1201 for N44 and N48.

 

[adroc_thurston]

So FSR 4 with AI upscaling should have back compat to RDNA 3 but may not be effective on RDNA 2 ??

It's not a VRF size thing.
Anything with WMMA should work.

 

[gaav87]

Iv just heard Navi48 599$ slower than 4070ti super

 

[Tuna-Fish]

Iv just heard Navi48 599$ slower than 4070ti super

Heard from where?

 

[gaav87]

Heard from where?

Not from mlid thats for sure. You will see in 3weeks.

 

[Win2012R2]

just heard Navi48 599$

Battlemade made it very unlikely, but one can always count on AMD doing very dumb pricing at start

 

[jpiniero]

Iv just heard Navi48 599$ slower than 4070ti super

5070 would have to be more than $599... which it probably is.

 

[beginner99]

Iv just heard Navi48 599$ slower than 4070ti super

That would be worse than what I imagined even in a realistic scenario. I mean the performance was clear will not be great simply from the die size. that isn't really a surprise. but $599 is just terrible. that's not even mainstream in any way or fashion. And since the die is small and yields high, AMD would what as many on the full chip as possible so the cut version will likley be the same stupid thing as with the 7700 vs 7800.

GPUs don't really age. that pricing would be in competition with 5070 and used 4070 ti supers and especially vs the later not having vra, advantage. given the pro nvidia bias, that won't end well. if this really happens, AMD should just close the GPU division. not even trying is pointless. Either go for the grown or market share but just selling small dies at insane prices, why? what is the strategy here? I don't see it.

 

[marees]

Heard from where?

AIB sources:
GB205-300: $699
NavI 48 XTX: $599

https://twitter.com/x/status/1870358327806312727

 

[marees]

AIB sources:
GB205-300: $699
NavI 48 XTX: $599

https://twitter.com/x/status/1870358327806312727

Slightly cheaper 7900xt with better RT (This is in line with my expectations)

https://twitter.com/x/status/1870452660345917832

 

[SolidQ]

Slightly cheaper 7900xt with better RT

how it can be slower than 4070ti super, if 7900xt a bit faster, than super.
Maybe they include Wukong and Alan Wake 2 PT, which heavily pro NV games

 

[Tuna-Fish]

I wonder how they will price the rumored cut 16GB version.

 

[marees]

I wonder how they will price the rumored cut 16GB version.

My guess:

$600, $500 for 16gb n48
$400 for 12gb n48

$350 for 16gb n44
$300, $250 for 8gb n44

In light of competition from battlemage at low end & blackwell at high end, I believe Radeon's line-up will look like the below:

$200 — 7600 (6nm)
$250 — 7600xt / 28 CU 8gb navi 44xl
$300 — 32 CU 8gb navi 44
$350 — 7700xt / 16gb 32 CU navi 44xt
$400 — 48 CU 12gb navi 48xl
$450 — 7800xt
$500 — 7900gre / 56 CU 16gb navi 48xt
$600 — 64 CU 16gb navi 48xtx
$650 — 7900xt
$750 — 7900xtx / 32gb 64 CU navi 48xtx

 

[Saylick]

Slightly cheaper 7900xt with better RT (This is in line with my expectations)

https://twitter.com/x/status/1870452660345917832

According to TPU, the 7900XT is on par with a 4070 Ti Super so….

 

[Timorous]

No chance, they even have a 96-bit SKU planned.

96bit with 12GB clamshell would be a lot better than 128bit with 8GB of Vram imo.

 

[GodisanAtheist]

This isn't the future I was sold in my childhood and I want a refund.

We're supposed to have flying cars and holodecks God damn it, not $600 270mm2 dies!

 

[Kepler_L2]

not $600 270mm2 dies!

Evil AMD making $400 of profit per card according to some people here

 

[ajsdkflsdjfio]

Evil AMD making $400 of profit per card according to some people here

If it were on TSMC N5 and used rx 7600 AIB parts it could be!! I wouldn't call it profit though but raw difference between production costs and sales cost.

202 mm2 RX 7600 die is around 35$ on a 9725$ N7 wafer, 272 mm2 die on 12730$ wafer is around 65$ accounting for 85% yield. However RDNA4 is using TSMC 4nm so.

If RX 7600 profit margins are 25% at 250$, a Navi48 card using typical RX7600 components would still be in the "profit" at 250$. But this is minus all other associated costs with designing and selling GPUs.

Its less about die size and more about things like components, RAM, AIB margins, parametric yield etc. If only product prices were as simple as die costs though LOL, we could have a 200$ 7900xtx or a 4090 for 400$! No marketing costs, R&D, taxes, shipping costs, board partner costs. What a simple world it would be...

But alas we live in the real world, so yes a 400 dollar difference between the raw costs associated with producing a card and the actual sale price of it isn't that insane. But to call it "profit" is very misleading.

I mean a 4090 die yields around 90 chips per wafer, let's say Jensen pays a cool 15k for an 4N TSMC wafer. Let's also assume a 50% yield rate per wafer. That'd mean 45 dies for 15k or around 333.33 dollars per die, and that's being extremely charitable in both low yields and not much of a discount for wafer prices. 333.33 die cost versus a 2k street price? With the rest of the components being 400$, the so called "profit" is above 1300$. And this is assuming an abysmal 50% yield, pretty much no wafer discount, and all the defective dies being thrown away instead of cut down and repackaged into another 1k+ card.

This u btw?

 

[SolidQ]

8xxx ref?

 

[Win2012R2]

I wouldn't call it profit though but raw difference between production costs and sales cost.

It's called "gross profit"

 

[SolidQ]

AIB sources:
GB205-300: $699
NavI 48 XTX: $599

That mean 5070 is faster or just for green label and 12vram extra 100$

 

[ajsdkflsdjfio]

It's called "gross profit"

Right, so once again how is a 350$ to 400$ gross profit on a Navi48 card using all the same parts as a rx7600 minus the die?

BTW Intel gross profit was 43 bil in 2021 vs a revenue of 79 bil. And that includes marketing and sales costs. Almost like the actual costs of producing the products is only half of the equation or even less!!

 

[Win2012R2]

Right, so once again how is a 350$ to 400$ gross profit on a card somehow insane?

It's insane to suggest that AMD makes so much now on 270 sq mm consumer GPU.

 

[ajsdkflsdjfio]

It's insane to suggest that AMD makes so much now on 270 sq mm consumer GPU.

Gross profit or net income? Factor in all other costs and you might be only making 100 dollars from a gross profit of 400 dollars. Just enough to pay for R&D.

Like I said, if the rx 7600 is making 25% gross profit at 250$, Navi 48 can absolutely be making 350-400$ gross profit selling at 600$ with only a +30$ die using N5. However it's likely to be less than that because navi48 cards are likely to be larger, have more RAM, and use a more expensive PCB. So let's say 250-300$ gross profit.

Saying 270mm2 die means nothing... because the die itself is barely +40-50 dollars MAX on N4.

To imply that a 270mm2 die cannot make 400$ gross profit is dumb asf anyways. The 4070 ti for example is a 294 mm^2 die size. The die itself likely costs somewhere around 75$ even assuming no significant TSMC discounts for Nvidia. With an MSRP of 800 dollars, unless the rest of the components are somehow significantly more than 330$(which is already an insane figure), the 4070 ti gross profit is ATLEAST 400$.