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

 

[soresu]

unless there is a UDNA Navi 53 waiting in the wings that can do it

UDNA will not be codenamed Navi going by what Kepler has said.

Probably Sirius based on this image would be my guess....

 

[poke01]

In the way that ROCm while still very flawed has come a long way since the last time the AMD GPU ecosystem was unified with Vega.

Well at this point its all hardware, hopefully RDNA4 RT/AI cores provide a good start

 

[ToTTenTranz]

Simply that they have put all pedals to the metal on UDNA roadmap.

The problem is they cut the RDNA4 lineup down to only 2 chips, supposedly to speed up the roadmaps. Yet what ended up happening was RDNA4 is only releasing in 2025 (meaning it probably got delayed a bunch of quarters), and UDNA is only releasing in 2026.


So what seems to have happened is we're getting all the downsides of cutting down the RDNA4 lineup (delayed release, no halo product, loss of mindshare, loss of competitiveness and Nvidia charging as much as they want on the high end) while not getting any upside at all because UDNA is still only slated to 2026.

And from the looks of it, UDNA is only coming in late 2026 (assuming it doesn't delay into 2027) so AMD will once again go almost 2 years without a new architecture and GPU lineup.

 

[marees]

The problem is they cut the RDNA4 lineup down to only 2 chips, supposedly to speed up the roadmaps. Yet what ended up happening was RDNA4 is only releasing in 2025 (meaning it probably got delayed a bunch of quarters), and UDNA is only releasing in 2026.


So what seems to have happened is we're getting all the downsides of cutting down the RDNA4 lineup (delayed release, no halo product, loss of mindshare, loss of competitiveness and Nvidia charging as much as they want on the high end) while not getting any upside at all because UDNA is still only slated to 2026.

And from the looks of it, UDNA is only coming in late 2026 (assuming it doesn't delay into 2027) so AMD will once again go almost 2 years without a new architecture and GPU lineup.

Inexplicable if they don't have a navi 31 replacement until 2027. That is almost PS6 time

I am hoping it will be there by 2025 end rather (aligning with xbox next)

 

[el etro]

AMD needs to improve the performance per teraflop in its GPUs.

 

[SolidQ]

The problem is they cut the RDNA4 lineup down to only 2 chips

That my thoughts also, i think they decide spend more money on new arch, and reduce RDNA4

 

[blckgrffn]

I think RDNA 4 being cut down to size was a combination of

1. UDNA ambitions
2. TSMC (packaging) constraints
3. RDNA 3 being a misfire

All of this still doesn't explain why we don't have a monolithic Navi 49 that replaces Navi 31, unless there is a UDNA Navi 53 waiting in the wings that can do it

Signs point to RDNA4 being more of a og RDNA type release where we saw only the 5700XT and down and RDNA 2 showed their real ambitions/vision.

While it (RDNA1) was a big step over Vega it wasn’t “worth” having the upper versions of if they that would have delayed RDNA2 in the slightest.

Even more critical if UDNA is bringing a better AI feature set to all of their product lines, including Server & Professional.

I too want RDNA4 out soon, though. Let’s move on already!

 

[GodisanAtheist]

The problem is they cut the RDNA4 lineup down to only 2 chips, supposedly to speed up the roadmaps. Yet what ended up happening was RDNA4 is only releasing in 2025 (meaning it probably got delayed a bunch of quarters), and UDNA is only releasing in 2026.


So what seems to have happened is we're getting all the downsides of cutting down the RDNA4 lineup (delayed release, no halo product, loss of mindshare, loss of competitiveness and Nvidia charging as much as they want on the high end) while not getting any upside at all because UDNA is still only slated to 2026.

And from the looks of it, UDNA is only coming in late 2026 (assuming it doesn't delay into 2027) so AMD will once again go almost 2 years without a new architecture and GPU lineup.

-I figure that the Hardware is done, validated, and ready to rock but the software (not simply drivers, but maybe FSR 4.0) just isn't ready to roll yet.

AMD has realized a day late and dollar short as always that if you're not launching your new GPU hardware with a new GPU feature then no one is buying (esp if you're missing the Halo and Enthusiast segments).

 

[blckgrffn]

-I figure that the Hardware is done, validated, and ready to rock but the software (not simply drivers, but maybe FSR 4.0) just isn't ready to roll yet.

AMD has realized a day late and dollar short as always that if you're not launching your new GPU hardware with a new GPU feature then no one is buying (esp if you're missing the Halo and Enthusiast segments).

Oh man, that would destroy my AMD experience of getting a fresh new card and then dealing with mystery issues for at least a couple of months!

If true, the "best" part for them and everyone else would be that the containers of cards would be on the water right now - or already landed stateside - which means that they would have a solid launch free of any political shenanigans that might be coming in Q1/Q2.

I want one, so I am projecting a bit.

 

[igor_kavinski]

Inexplicable if they don't have a navi 31 replacement until 2027. That is almost PS6 time

Makes it sound almost like AMD wants to veer gamers towards consoles...

 

[marees]

"at least" one new Radeon RX 8000 series graphics card hitting the market in late January.

Read more: https://www.tweaktown.com/news/1018...one-new-radeon-gpu-in-late-january/index.html


this source said they couldn't confirm if AMD will have its next-gen FSR 4 upscaling technology ready for RDNA 4 and CES 2025.

 

[gdansk]

MLID sourced garbage article

 

[burninatortech4]

MLID sourced garbage article

MLID being used as fantasy conjecture can be entertaining in small doses - but using it as a "source" is pretty brain dead. Agreed.

 

[Mopetar]

UDNA will not be codenamed Navi going by what Kepler has said.

Probably Sirius based on this image would be my guess....

I for one am glad to hear that AMD is finally getting Sirius with their GPUs.

 

[soresu]

-I figure that the Hardware is done, validated, and ready to rock but the software (not simply drivers, but maybe FSR 4.0) just isn't ready to roll yet.

Nah it's definitely RDNA3 channel volume they want down to a minimum.

 

[soresu]

I for one am glad to hear that AMD is finally getting Sirius with their GPUs.

Yes I'm Sirius, and don't call me Shirley.....

 

[el etro]

dunno what silicon process AMD will use.

 

[el etro]

I would love if AMD bump up RT performance if is said on this thread is real

 

[igor_kavinski]

ROCm 6.3 adds several new features including a Fortran compiler, and SGLang

ROCm gets more abilities for enterprise customers to take advantage of

www.tomshardware.com

Fortran compiler! Encroaching into Intel's territory.

 

[maddie]

ROCm 6.3 adds several new features including a Fortran compiler, and SGLang

ROCm gets more abilities for enterprise customers to take advantage of

www.tomshardware.com

Fortran compiler! Encroaching into Intel's territory.

Wow, Fortran, my 1st language, decades old memories.

 

[igor_kavinski]

Wow, Fortran, my 1st language, decades old memories.

Now you can relive them if you have a Radeon 6000 at least

 

[Timorous]

I am so bored. Where is the false hype of a 5,000 GHz 1,024 WGP 128GB 7090 killer released q1 next year for $400 so we can do silly stuff like make up product stacks based on those outrageous claims and then be sad when AMD don't meet those expectations?

 

[GodisanAtheist]

I am so bored. Where is the false hype of a 5,000 GHz 1,024 WGP 128GB 7090 killer released q1 next year for $400 so we can do silly stuff like make up product stacks based on those outrageous claims and then be sad when AMD don't meet those expectations?

- Nah guy its going to have 42 GCDs and the answer to everything.

 

[Timorous]

- Nah guy its going to have 42 GCDs and the answer to everything.

Think bigger, 42 GCDs, 42 TCDs on a single card but AMD have perfected xfire so they can sell you a 42 card stacked unit which can go into a 42 cabinet cluster. 42's all the way down.

 

[MrTeal]

Think bigger, 42 GCDs, 42 TCDs on a single card but AMD have perfected xfire so they can sell you a 42 card stacked unit which can go into a 42 cabinet cluster. 42's all the way down.

42GB VRAM on a 448bit memory bus?