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

techjunkie123 · Jan 30, 2025

Meteor Late said:
So if it's 345-350 mm2, we know RTX 5080 is 378 mm2, if 9070 XT is like 8-10% slower than 5080, seeing as 5080 is around 10% faster than 4080 super and 9070 XT will have around that level of performance. It means PPA is similar between AMD and Nvidia on the same node WITH worse memory.

Keep in mind that's only raster performance. RT performance will be at least one class behind, maybe between a 5070 and 5070 Ti. So if that performance is true then slightly worse PPA.

lightmanek · Jan 30, 2025

adroc_thurston said:
These are words and no they're not gonna be aggressive with their pricing. No one buys Radeon.

Wrong!
I buy Radeons, ever since they were called ATi Rage IIc ...

I'm a minority, but saying no one buys them is wrong

igor_kavinski · Jan 30, 2025

SK10H said:
I test it on 7900xtx but with 24.12.1 driver on LM studio 0.3.8 on Win10, using Deepseek R1 distill owen 14B/32B. It outputs Gibberish no matter how simple the question.

Experience the DeepSeek R1 Distilled 'Reasoning' Models on AMD Ryzen™ AI and Radeon™

Reasoning models are a new class of large language models (LLMs) designed to tackle highly complex tasks by employing chain-of-thought (CoT) reasoning with the tradeoff of taking longer to respond. The DeepSeek R1 is a recently released frontier “reasoning” model which has been distilled into...

community.amd.com

Try Step 1 again

JustViewing · Jan 30, 2025

I think AMD missed a big free marketing opportunity with the delay of RDNA4. I would assume RDNA4 would have run circles around 7900XTX running local LLMs.

adroc_thurston · Jan 30, 2025

JustViewing said:
I would assume RDNA4 would have run circles around 7900XTX running local LLMs

Oh no. Stuff that no one cares about.

SK10H · Jan 30, 2025

igor_kavinski said:
Experience the DeepSeek R1 Distilled 'Reasoning' Models on AMD Ryzen™ AI and Radeon™

Reasoning models are a new class of large language models (LLMs) designed to tackle highly complex tasks by employing chain-of-thought (CoT) reasoning with the tradeoff of taking longer to respond. The DeepSeek R1 is a recently released frontier “reasoning” model which has been distilled into...

community.amd.com

View attachment 115920

Try Step 1 again

I will stick with ollama since it's open sourced as well unlike LM studio, but setup took a bit longer.
Also, last year the monthly drivers weren't that good, kept having stuttering issue, 24.12.1 works fine for my games, last one was 24.3.1, so I will stick with it for a while, this is just quick play on the side to exploit more value from the card. 🙂

Tup3x · Jan 30, 2025

lightmanek said:
I buy Radeons, ever since they were called ATi Rage IIc ...

I had such a bad experience with that particular card that I still hold grudges.

Meteor Late · Jan 30, 2025

Is there any info about the improved encoders in RDNA4? did AMD finally include 4:4:4 hardware encode and decode support?

lightmanek · Jan 30, 2025

Tup3x said:
I had such a bad experience with that particular card that I still hold grudges.

For me, it was cheap and cheerful and did output 800x600 in 16bits to my VGA, soon it was replaced with S3 Virge 3D, which was not a step in the right direction, followed by Riva 128 4MB.
Good old times when driver update could give 50% more FPS in a game or ruin another title ... loved it! (That's why I bought S3 Savage 3D after Riva 128ZX )

tajoh111 · Friday at 12:22 AM

H T C said:
AMD has a "golden opportunity" here to REALLY mess up nVidia, now that "the cat is out of the bag" and the 5090 is not as powerful (relative to the 4090) as they feared.

It OBVIOUSLY depends on the performance of their cards BUT if AMD's top GPU is:

- better than the 5080 (VERY unlikely), then price it @ ... say ... $600
- a bit worse than the 5080 (only a bit), then price it @ ... say ...$500
- considerably worse than the 5080 (likely), then price it @ ... say ... $400

UNLESS they're THIS aggressive with their pricing, i don't see AMD having better luck with this generation than they had in their current one.

OBVIOUSLY, this is gonna "wreck havoc" with the current cards still up for sale and that will cost A TON of money: it's SUPPOSED to be a "calculated risk", designed to MASSIVELY increase their market share.

If it works, they'll end up losing QUITE A BIT of money NOW and earn A LOT of money later ... but if it doesn't ... they'll lose A LOT of money now AND later ...

The thing if AMD tries to get into a price war with Nvidia, they will lose.

Nvidia if they get beaten enough will drop the price substantially or release a new SKU to compete. They have done it in the past and they can obviously afford it more than ever.

Getting into a price war this generation poisons the videocard market for next generation while killing the profits on the current generation. It is a lose lose situation. What I means is by giving too much of a price to performance increase this generation AMD will be forced to use a large die to get a meaningful performance increase(since increases with nodes are not as good) while giving people high expectation that these cards should be relatively cheap because last gens was cheap. With 3nm chips and R and D being so expensive, this would sabotage both companies and would be pyrrhic victory for AMD financially.

reaperrr3 · Friday at 12:31 AM

StinkyPinky said:
GeRadeon

Raydeon (stronger raytracing focus, now that they're gonna be more competitive there with RDNA4)?
Or Rydeon, ryding the success of Ryzen

marees · Friday at 1:03 AM

9070xt to match 4070 ti super on RT
9070 to match 4070 super on RT

Pure raster, 9070xt to fall short of 7900 xtx

adroc_thurston · Friday at 1:05 AM

marees said:
Pure raster, 9070xt to fall short of 7900 xtx

they're pretty much the same.
It's a good part they can't truly sell.

SolidQ · Friday at 1:10 AM

adroc_thurston said:
they're pretty much the same

RT for real 4070S and 4070tiS? Because i know usually after release new gen AMD add like 10-15% perf to RT

adroc_thurston · Friday at 1:15 AM

SolidQ said:
RT for real 4070S and 4070tiS?

a) no idea
b) don't care about FF mysterymeat
c) the chinese guy with N48 board on hand posted monke and CP2077 numbers anyway

del42sa · Friday at 2:07 AM

StinkyPinky said:
GeRadeon

GreeDeon 👍

coercitiv · Friday at 3:54 AM

Folks, stop trying to come up with rebranding option better than Amedeon. Won't happen.

gaav87 · Friday at 5:10 AM

Lets say n4p is 16000$ per 300mm wafer
And n4p is 13x27 ~351mm2 this means 160 dies per wafer
80% yield = 128 dies per wafer = 125$ per n48 (without dicing, testing, packaging, and binning)
Packaged and binned N48 ~150$
16GB gddr6 = spot pricing~36$
My total BOM est. is less than 300$
Lets say 300$ 1.5x gross profit margin = 450$ (chinese chiphell amd insiders still saying yesterday 479$ was the price shortly after ces)
(If the 300mm wafer is 20k = 500$, 24k = 525$)
So even at 24k$ per wafer and 80% gross margin 630$

SteinFG · Friday at 5:19 AM

gaav87 said:
Lets say 300$ 1.5x gross profit margin = 450$

That's gross margin of 33%, not 50%. (450-300)/450 is 33%

gaav87 said:
So even at 24k$ per wafer and 80% gross margin 630$

80% gross margin means only 20% is actual BOM. the GPU cost with 80% GM would be like $1700 😂

gaav87 · Friday at 5:23 AM

SteinFG said:
That's gross margin of 33%, not 50%. (450-300)/450 is 33%

80% gross margin means only 20% is actual BOM. the GPU cost with 80% GM would be like $1700 😂

i ment net margin i have fever...
Anyway if they price it at 649$ and the total bom is 350$ gros margins will be close to 50%

Tuna-Fish · Friday at 5:55 AM

gaav87 said:
i ment net margin i have fever...

No margin is ever computed on top of the cost like that. It's always % of the price.

gaav87 · Friday at 6:13 AM

SteinFG said:
That's gross margin of 33%, not 50%. (450-300)/450 is 33%

80% gross margin means only 20% is actual BOM. the GPU cost with 80% GM would be like $1700 😂

Tuna-Fish said:
No margin is ever computed on top of the cost like that. It's always % of the price.

Fixed looks good now ?

Lets say the cost for an N4P 300mm wafer is 16k$. And 9070xt die is ~13mm x 27mm =~351mm² this gives ~160 dies per 300mm wafer

Assuming an 80% yield, that gives us:
160 dies * 80% = 128 dies per wafer
Cost per die before dicing, testing, packaging, and binning = 16k$ / 128 = $125 per die

After packaging and binning, let's estimate the cost per N48 die at $150.

For additional components:
16GB GDDR6 memory at current spot pricing is ~$36.

This brings the total Bill of Materials (BOM) estimate to:
$150 (N48 die) + $36 (memory) = $186. Let's account for other minor components or underestimations and say the total BOM is 300$.

If we take the chiphell rumored price of $479 for 9070xt:
Gross profit margin = (($479 - $300) / $479) * 100 ≈ 37%
AMD's gross profit margin at the end of 2024 was reportedly ~48%.

To achieve a 49% gross margin:
At a BOM of $300, the 9070xt selling price would need to be ~$599 ($300 / (1 - 0.49) ≈ $599).

Taking different wafer costs:
If the 300mm wafer is 20k$:
BOM increases to $350 (assuming die costs scale linearly with wafer cost), leading to a price of $649 for a 49% margin.
If the 300mm wafer is 24k$:
BOM would be around $400, leading to a price of $699 for a 45% margin.

So even if a wafer costs 24k$, the total BOM would be $400 (if we maintain the original die cost calculation but adjust ONLY for the higher wafer price), and the card should cost $699 to achieve the 45% gross margin.

Note:
The calculations assume that die yield and other costs remain constant with changes in wafer cost, which IS NOT accurate in real-world scenarios.
The BOM includes only major components; other minor costs could push the actual BOM lower or slightly higher!

SteinFG · Friday at 6:45 AM

599 for MBA models is what I expect from AMD. Probably beats 5070 Ti in raster, no idea how good in RT (can get close, considering no RT gains in blackwell). 649-699 starting msrp would be so disappointing really

igor_kavinski · Friday at 6:57 AM

lightmanek said:
For me, it was cheap and cheerful and did output 800x600 in 16bits to my VGA, soon it was replaced with S3 Virge 3D, which was not a step in the right direction, followed by Riva 128 4MB.
Good old times when driver update could give 50% more FPS in a game or ruin another title ... loved it! (That's why I bought S3 Savage 3D after Riva 128ZX )

Thanks to The Tech Report and Anandtech, I never bought a loser card. And now both those sites are dead

gaav87 · Friday at 8:16 AM

Only some ML missing and DCC on 29th. They added DCC today so 9070xt linux drivers are good to go

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Member

Senior member

Lifer

Senior member

Diamond Member

Member

Golden Member

Senior member

Senior member

Senior member

Member

Senior member

Diamond Member

Golden Member

Diamond Member

Member

Diamond Member

Senior member

Senior member

Senior member

Golden Member

Senior member

Senior member

Lifer

Senior member