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

beginner99 · Jan 28, 2025

gaav87 said:
Chiphell boss just said after azor tweet price is now under 5k yuan for oc xt and 4k+ for ref so 499$-613$ ? China has 13% tax right i dont believe it.

Why? $499 and $599 in my opinion have always been sensible price points IF AMD wants to gain market share. since 9070 seems to be a lot lower clocked and likely power limited, it's possible also the XT is $619.

If AMD are happy with 10%, then yeah, expect $699 as XT will likley beat a 5070 ti in raster.

gaav87 · Jan 28, 2025

beginner99 said:
Why? $499 and $599 in my opinion have always been sensible price points IF AMD wants to gain market share. since 9070 seems to be a lot lower clocked and likely power limited, it's possible also the XT is $619.

If AMD are happy with 10%, then yeah, expect $699 as XT will likley beat a 5070 ti in raster.

No he ment 499$ for 9070xt reference and ~613$ for aib oc models. not 9070 (non-xt) for 499$. thats why i dont believe it 499$ for 9070xt would murder every card even 5080

adroc_thurston · Jan 28, 2025

gaav87 said:
thats why i dont believe it 499$ for 9070xt would murder every card even 5080

Kinda the point, they have the BOM leeway.

marees · Jan 28, 2025

FSR4 ML model is FP8-based.

Only RDNA 4 adds WMMA support for FP8.
Running on RDNA 3 will need some change. It supports BF16, might be close enough perf-wise;
same for XDNA 2 (Ryzen AI). Re-quantization to INT8 is another option.
For RDNA 2 and below, forget it.

https://twitter.com/x/status/1883889129894908258

Kepler_L2 · Jan 28, 2025

gaav87 said:
No he ment 499$ for 9070xt reference and ~613$ for aib oc models. not 9070 (non-xt) for 499$. thats why i dont believe it 499$ for 9070xt would murder every card even 5080

There is no reference card.

gaav87 · Jan 28, 2025

Kepler_L2 said:
There is no reference card.

Really ? This look like mba to me ?

Kepler_L2 · Jan 28, 2025

gaav87 said:
Really ? This look like mba to me ?
View attachment 115762 View attachment 115763

It won't be released.

MrTeal · Jan 28, 2025

Be interesting to see if there's actually any cards at MSRP then, or if it will be more like we're seeing with Blackwell where all the AIB's are eschewing MSRP cards.

gaav87 · Jan 28, 2025

Kepler_L2 said:
It won't be released.

Hotspot problem ?

blckgrffn · Jan 28, 2025

Kepler_L2 said:
It won't be released.

I am surprised they aren’t doing a white label (ODM?) card to say they are hitting whatever the “MSRP” is and only selling it direct and at a partner like Microcenter to allow them to say the price is x in all the reviews but it’s really x+AIB premium ala Nvidia.

I think it’s disingenuous as a practice but obviously it’s trick of the trade so I’d expect them to use it.

Win2012R2 · Jan 28, 2025

Can't AMD get a Kickstarter project for a Big Chip? Get firm pre-orders for half a bil that should cover costs...

lucasworais · Jan 28, 2025

Win2012R2 said:
Can't AMD get a Kickstarter project for a Big Chip? Get firm pre-orders for half a bil that should cover costs...

They have the money, they just don't have the balls.

coercitiv · Jan 28, 2025

Win2012R2 said:
Can't AMD get a Kickstarter project for a Big Chip? Get firm pre-orders for half a bil that should cover costs...

Let me rephrase this as a newspaper title:

200 Billion dollar company asks kids for their lunch money to develop AI chips.

Win2012R2 · Jan 28, 2025

lucasworais said:
They have the money, they just don't have the balls though.

That's why I proposed kick-starter...

blckgrffn · Jan 28, 2025

Win2012R2 said:
That's why I proposed kick-starter...

So the Kickstarter is acquire testicles, then?

ToTTenTranz · Jan 28, 2025

marees said:
FSR4 ML model is FP8-based.

Only RDNA 4 adds WMMA support for FP8.

Running on RDNA 3 will need some change. It supports BF16, might be close enough perf-wise;

same for XDNA 2 (Ryzen AI). Re-quantization to INT8 is another option.

For RDNA 2 and below, forget it.

https://twitter.com/x/status/1883889129894908258

I wonder if XDNA2 can realistically be used for an upscaler. How close is the NPU to the GPU?
I'm guessing there's no direct bus between each others' caches, so everything must go through the RAM.

uzzi38 · Jan 28, 2025

ToTTenTranz said:
I wonder if XDNA2 can realistically be used for an upscaler. How close is the NPU to the GPU?
I'm guessing there's no direct bus between each others' caches, so everything must go through the RAM.

I don't think putting extra stress on the memory controller as you pass data from the GPU to the NPU and back again (going through main memory in both cases) is a very sensible idea for an already memory-bound APU...

And I also don't think the NPU is fast enough to warrant it.

Something like Microsoft's AutoSR can work because it doesn't pass data back to the iGPU for further processing afterwards, just to the display controller to output directly. But taht can only ever work for a spatial upscaler, not a temporal one.

branch_suggestion · Jan 29, 2025

N4C would've given NV R300 flashbacks.
Instead it is RV770 flashbacks, good enough I guess.

SolidQ · Jan 29, 2025

after seeing 5080 results, i think 9070vanilla gonna be faster, than 5070 vanilla

Kepler_L2 · Jan 29, 2025

SolidQ said:
after seeing 5080 results, i think 9070vanilla gonna be faster, than 5070 vanilla

RDNA4 is faster than Blackwell at equal SM/CU count, and 9070 is 56 CUs vs 48 SMs on 5070.

SolidQ · Jan 29, 2025

Kepler_L2 said:
RDNA4 is faster than Blackwell

Now is question how much it faster.
9070XT 64CU vs 5070ti 70sm, do you think 64CU > 70SM?

adroc_thurston · Jan 29, 2025

SolidQ said:
do you think 64CU > 70SM?

Yeah lol cuz it's also like 500MHz more in fmax.
Goddamit Lisa, this is your first big L in CEO tenure.

SolidQ · Jan 29, 2025

adroc_thurston said:
Goddamit Lisa, this is your first big L in CEO tenure.

Yeah it's, but also maybe someone will discuss inside company with Lisa and rethink strategy? I think with UDNA it's late, but with UDNA2 possible

adroc_thurston · Jan 29, 2025

SolidQ said:
I think with UDNA it's late, but with UDNA2 possible

It's not late even with RDNA4.
They only have to commit.

branch_suggestion · Jan 29, 2025

If N4C was really killed that late, they could totally get it out in time for Computex 2026.
Would be odd timing as RDNA5 would be out not long after, but both sides have done that kinda thing before.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Senior member

Diamond Member

Senior member

Senior member

Senior member

Senior member

Diamond Member

Senior member

Diamond Member

Senior member

Member

Diamond Member

200 Billion dollar company asks kids for their lunch money to develop AI chips.​

Senior member

Diamond Member

Senior member

Platinum Member

Senior member

Golden Member

Senior member

Golden Member

Diamond Member

Golden Member

Diamond Member

Senior member

200 Billion dollar company asks kids for their lunch money to develop AI chips.