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

adroc_thurston · Jan 14, 2025

Novacius said:
How many games are there with FSR 3.1 again versus games with DLSS 2.0 at least?

A lot? It's a pretty ubiquitous upscaling.

Keller_TT · Jan 14, 2025

adroc_thurston said:
no it's normal.
It's more like n31 is ass.

That's why I'm skeptical. This means they've recovered the lost performance of RDNA 3 and did another impressive step and with negligible errata.
RTRT, I do expect that leap because of what's going in and how much was missed previously, but raster, I'll wait and see.

PJVol · Jan 14, 2025

adroc_thurston said:
What does "+10% at best" mean.

+10% to 7900xtx

adroc_thurston · Jan 14, 2025

PJVol said:
+10% to 7900xtx

ah. no.
Also that's not a singular workload.

Keller_TT said:
This means they've recovered the lost performance of RDNA 3 and did another impressive step and with negligible errata.

yeah that's called executing on your own roadmap.
I know that people (well, plebbit) crave anudda Vega from AMD GPU-wise, always, forever, to justify green slop purchase but it ain't happening.

techjunkie123 · Jan 14, 2025

Novacius said:
Stop comparing it to the previous generation and look at the 5070 instead. It's very possible the 9070 XT is 20-30% faster in raster and just about as fast in RT. Nvidia on the other hand has a feature advantage, especialy with DLSS. With DLSS 4 it will be possible to override the DLSS version via driver in every game that has at least DLSS 2.0, so a lot of games from the past five years, which is massive. Also Nvidia has mindshare and is often viewed as the default choice. To overcome that and get more marketshare, AMD can't price the 9070 XT even the same as the 5070. $50 difference is the minimum and gets AMD below the $500 barrier, which is psychologically important I think, even if the difference isn't that big.

Sure. +25% raster, say comparable to the 5070 Ti, and RT ~ 5070 even in heavy RT scenarios which means a win in light RT cases.

Price that at 550$ at the most. Previously AMD had to price the 7800XT cheaper than the 4070 because it lost in RT. If that changes this gen, I think they can price the same. Esp with 16 GB VRAM vs 12 GB.

Keller_TT · Jan 14, 2025

techjunkie123 said:
Bruh we're talking raster performance that previously cost 900$+ and RT performance that cost at least 750$+. Now being sold for 600$ (??), using a similar die size. How is that unreasonable??

I don't get the hype that expects AMD to price the 9070XT at 500$ or less. (If performance is indeed 4080 raster and 4070 Ti Super in RT)

That's how silicon technology and gen on gen progression works. Nvidia reference pricing has gone without limits as they have no competition.
RDNA2 to RDNA4 should be over 1.8x the density for a modest wafer cost rise at worst.
N48 die size is quite close to a 7800 XT and should be cheaper to make than N32 due to design costs, DDR6 memory, etc. I made a post few days ago with a rough calculation and modest PPW figures.
It just means what they hugely missed with RDNA3 and the market dunked them, they'll be rectifying now.

But before all this, they have to deliver on performance as well as efficiency. There's another aspect regarding software support and performance. Outside gaming, AMD is really bad for design & engineering. I always joke that AMD has to be doin design and simulation of their cards using Nvidia cards.

adroc_thurston · Jan 14, 2025

Keller_TT said:
I always joke that AMD has to be doin design and simulation of their cards using Nvidia cards.

EDA stuff runs exclusively on CPUs. X3D CPUs, really.

Keller_TT · Jan 14, 2025

adroc_thurston said:
ah. no.
Also that's not a singular workload.

I guess the OP asked about a +10% RTRT performance increase over 7900 XTX. Not raster. Normally, it should be if baseline raster performance is that much better and CUs can do RT much better in RDNA4.
The leaks of the AIB XT card placed it at more than 7900XTX + 10% in heavy RT. Nudging the 4070 TiS.

adroc_thurston · Jan 14, 2025

Keller_TT said:
Normally, it should be if baseline raster performance is that much better and CUs can do RT much better in RDNA4.

Raster is trivial to eval since the nature of it hasn't majorly changed since DX11.
RTRT is DX7-era mysterymeat FF blobs.

Keller_TT said:
I guess the OP asked about a +10% RTRT performance increase over 7900 XTX

It's mysterymeat FF workload, the gains will be all over the place depending on how well stuff maps to the new FF block and the new caching.

adroc_thurston · Jan 14, 2025

xpea said:
hmmm hmmm

Major EDA vendors utilizing GPU acceleration:
Companies like Synopsys, Cadence, Ansys, Siemens and Mentor Graphics have integrated GPU acceleration into their EDA suites, including tools like circuit simulators, place and route engines, and static timing analyzers.

Examples of GPU-accelerated EDA tasks:

Circuit simulation:
Running SPICE simulations on a GPU can drastically reduce simulation time for analog and mixed-signal designs.

Placement and routing:
Optimizing the layout of components on a chip by exploring various placement options with parallel processing on the GPU.

Static timing analysis:
Analyzing the timing constraints of a circuit to ensure proper functionality, significantly accelerated with GPU computing

More reading:

Nvidia’s role in the EDA industry - EDACafe Editorial

Not just GPU-based acceleration: the partnerships announced on occasion of this year’s GTC event demonstrate that Nvidia software, too, is a key technology in several existing or upcoming EDA tools EDA and engineering software received quite a bit of attention at the recent Nvidia GTC event...

www10.edacafe.com

Synopsys Showcases EDA Performance and Next-Gen Capabilities with NVIDIA Accelerated Computing, Generative AI and Omniverse

Expanding Decades of Collaboration Across Synopsys' Full EDA Stack News Highlights: Enabling Synopsys' industry-leading full AI-driven EDA stack on the NVIDIA...

www.prnewswire.com

You should really go back to plebbit (or b3d) instead of trying to akshually me.
Running EDA stuff on SIMD machines is still an utterly memeworthy endeavor.
Maybe less so in 5 years, but now? Lmao.

Hitman928 · Jan 14, 2025

xpea said:
hmmm hmmm

Major EDA vendors utilizing GPU acceleration:
Companies like Synopsys, Cadence, Ansys, Siemens and Mentor Graphics have integrated GPU acceleration into their EDA suites, including tools like circuit simulators, place and route engines, and static timing analyzers.

Examples of GPU-accelerated EDA tasks:

Circuit simulation:
Running SPICE simulations on a GPU can drastically reduce simulation time for analog and mixed-signal designs.

Placement and routing:
Optimizing the layout of components on a chip by exploring various placement options with parallel processing on the GPU.

Static timing analysis:
Analyzing the timing constraints of a circuit to ensure proper functionality, significantly accelerated with GPU computing

More reading:

Nvidia’s role in the EDA industry - EDACafe Editorial

Not just GPU-based acceleration: the partnerships announced on occasion of this year’s GTC event demonstrate that Nvidia software, too, is a key technology in several existing or upcoming EDA tools EDA and engineering software received quite a bit of attention at the recent Nvidia GTC event...

www10.edacafe.com

Synopsys Showcases EDA Performance and Next-Gen Capabilities with NVIDIA Accelerated Computing, Generative AI and Omniverse

Expanding Decades of Collaboration Across Synopsys' Full EDA Stack News Highlights: Enabling Synopsys' industry-leading full AI-driven EDA stack on the NVIDIA...

www.prnewswire.com

This is all stuff they’re working on in the IC design space, it’s not available today.

SolidQ · Jan 14, 2025

https://twitter.com/x/status/1879237767013769304

Better than expected?

adroc_thurston · Jan 14, 2025

SolidQ said:
https://twitter.com/x/status/1879237767013769304

Better than expected?

You're reading too much into random poasting.

gaav87 · Jan 14, 2025

SolidQ said:
https://twitter.com/x/status/1879237767013769304

Better than expected?

No he was thinking 9070xt and 9070 have the same cu count.

lucasworais · Jan 14, 2025

SolidQ said:
https://twitter.com/x/status/1879237767013769304

Better than expected?

9070 XT and non-XT have the same amount of shader cores

adroc_thurston · Jan 14, 2025

lucasworais said:
9070 XT and non-XT have the same amount of shader cores

they don't

blckgrffn · Jan 14, 2025

Mopetar said:
Although JHH would never do it, I'd wet myself laughing if NVidia also delayed to force AMD to launch their mid-tier first or delay even further.

Everything is hard launching launching Q2 now! In mid to late July!

(Probably)

poke01 · Jan 14, 2025

adroc_thurston said:
A lot? It's a pretty ubiquitous upscaling.

not as much as DLSS2.0
Also older games like RDR2 are still on FSR2 so those games cannot use FSR4 unless someone finds a way?
Maybe AMD will find a way

SolidQ · Jan 14, 2025

poke01 said:
Also older games like RDR2 are still on FSR2

game have already FSR 3.0

poke01 said:
cannot use FSR4 unless someone finds a way?

I think modders will fix it, like with FSR 3.1 mod

poke01 · Jan 14, 2025

SolidQ said:
game have already FSR 3.0

Not RDR2. RDR1 has FSR3.0

SolidQ said:
I think modders will fix it, like with FSR 3.1 mod

This is more likely

SolidQ · Jan 14, 2025

poke01 said:
Not RDR2. RDR1 has FSR3.0

oh forgot about RDR1 PC

Interesting will be DLSS same quality on switch 2 or that cut down version

adroc_thurston · Jan 14, 2025

poke01 said:
Also older games like RDR2 are still on FSR2 so those games cannot use FSR4 unless someone finds a way?

If they're just swapping the DLL then it's gucci?

Heartbreaker · Jan 14, 2025

poke01 said:
not as much as DLSS2.0
Also older games like RDR2 are still on FSR2 so those games cannot use FSR4 unless someone finds a way?
Maybe AMD will find a way

I expect the community will make a stub DLL, to convert from DLSS calls to FSR4 call. So any DLSS 2+ game should be convertible to FSR4 for scaling.

This already exist for converting DLSS games to to FSR. It probably won't take long until they get this working for FSR4.

FidelityFx Super Resolution 2.2 - FPS boost mod

Boost your FPS in Cyberpunk by replacing DLSS with FSR 2.2 which works on all GPUs.

www.nexusmods.com

There are also mods to replace DLSS FG with AMD FG...

eek2121 · Jan 14, 2025

Josh128 said:
So it seems, according to HXL on Xitter via Chiphell, that the yet unannounced RDNA 4 reveal / embargo lift is delayed again, to an undisclosed date. They say it is to let "Huang go first" again. Why though, if not to try and maximize pricing at the last minute? They already have 5000 series pricing and specs.

Seems they are back to their old trolling ways, thinking back to camping outside an Nvidia presentation with the RX 290X, and "Jebaiting" the price of the RX 5700XT, except even worse now, as they already have Nvidias pricing and specs.

??? Normally I do pay attention to those rumors, however, I don’t think that one is true. Reviewers already have cards and preorders are up.

adroc_thurston said:
no it's normal.
more like n31 is ass.

hell no, ATi did bigger, better and meaner things before.

they have to ship Navi40. Or any chungus Navi5c.

Ryzen wasn't a thing until 3950X, and that had no cooler.
Frankly Ryzen wasn't a thing until 5950X won it all.

I agree, though Zen 1 wasn’t terrible. I have no regrets about buying my 1950X. Yes, things didn’t really take off until later.

adroc_thurston said:
If they're just swapping the DLL then it's gucci?

No, the end result won’t be as great if they do that.

xpea · Jan 14, 2025

Nah.

Hitman928 said:
This is all stuff they’re working on in the IC design space, it’s not available today.

nope. being used by Nvidia for more than 4 years on cadence and Synopsys

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Member

Senior member

Diamond Member

Member

Member

Diamond Member

Member

Diamond Member

Diamond Member

Diamond Member

Golden Member

Diamond Member

Senior member

Member

Diamond Member

Diamond Member

Diamond Member

Golden Member

Diamond Member

Golden Member

Diamond Member

Diamond Member

Diamond Member

Senior member