Discussion RDNA 5 / UDNA (CDNA Next) speculation

[gdansk]

View attachment 129736
hell yeah time for the paint-off

Using the rectangle tool? You philistine.
You gotta free hand those.

 

[adroc_thurston]

Using the rectangle tool? You philistine.
You gotta free hand those.

pls no more kicking, I was drawing datboi while playing DoW: DE.

 

[Saylick]

View attachment 129736
hell yeah time for the paint-off

Inb4 Wtftech runs with this sketch.

 

[igor_kavinski]

I fed it into ChatGPT to clean it up and this is what it gave me lol

View attachment 129735

To Hassan, with love

 

[MrMPFR]

Reading about NVIDIA Volta's architectural changes but some of it is well above my head. The independent thread scheduling and convergence optimizer sounds like a big deal, but is this even practical outside of niche HPC and compute workloads?
Def can't see a relation to the GPU work graphs, but perhaps I'm wrong.

And more importantly does AMD have something like this rn and if not would it make sense to include it in future CDNA and RDNA, for example CDNA5 and RDNA5?

Edit: Yes It's still me Magras00, just realigning with name on other sites.

 

[adroc_thurston]

Reading about NVIDIA Volta's architectural changes but some of it is well above my head. The independent thread scheduling and convergence optimizer sounds like a big deal, but is this even practical outside of niche HPC and compute workloads?
Def can't see a relation to the GPU work graphs, but perhaps I'm wrong.

And more importantly does AMD have something like this rn and if not would it make sense to include it in future CDNA and RDNA, for example CDNA5 and RDNA5?

Forget about applying anything Nvidia to AMD GPUs.

 

[Mopetar]

Inb4 Wtftech runs with this sketch.

That's small potatoes. AMD should take it and use it in their official slides.

 

[itsmydamnation]

Creative actually got bigger after they left the GPU market after the Geforce 4 and focused on sound cards using their proprietary EAX sound processors. But shortly after that Microsoft blocked external sound APIs from DirectX in games, so for the past 2 decades they've been steadily but very slowly withering away. I think their founder died last year.
IMO they still make excellent sound products with great value. I use their Aurvana Ace 2 as my daily IEMs and they're pretty awesome. I still used their Gigaworks S750 7.1 in my desktop PC until a couple weeks ago (got tired of all the wiring and some of the speaker opamps finally started giving away after >20 years).


Diamond got sold to S3, and then S3 got sold to VIA, which was sold to HTC, and now somehow the brand belongs to TUL (Powercolor & Sparkle). I guess part of me wants to believe some of Diamond's engineers are now still designing graphics cards.

yes i had a creative 4x cdrom + SB16 combo in my 486-DX2-66.

I also still have PICE sound cards because i use ASIO for VST /AMP modelling so its more creative GPUs and Diamond as GPU manufactures. Also i happened to have a Diamond TNT2 ultra but in the era of GeForce 256 so that made me a poor peasant .


edit: maybe me google for nostalgia

Leadtek WinFast GeForce 256 DDR Review

Breaking free of the memory bottleneck chains.

www.tomshardware.com

shogo mobile armoured division .... man i loved that game.

 

[Claudiovict]

Leaker claims AMD RDNA 5 is bigger than expected - Larger CUs! - OC3D

A leaker has claimed that AMD has reworked its GPU Compute Units with RDNA 5, making AMD's next-gen GPUs may be a lot bigger than expected.

overclock3d.net

"AMD’s RDNA 5 GPUs could be much bigger than expected. According to the leaker zhangzhonghao, AMD has changed the structure of its Compute Units (CU) with RDNA 5. Instead of having 64 Stream Multiprocessors (SM) per Compute Unit, RDNA 5 reportedly features 128. That’s a 2x increase in SM count per CU."

I'm really new to these more indepth topics, this has something to do with the WGP vs CU previous talk? This 2x is really likely to hapen?

 

[ToTTenTranz]

I'm really new to these more indepth topics, this has something to do with the WGP vs CU previous talk? This 2x is really likely to hapen?

Yes, CUs are 2x wider now. A RDNA5 CU has the same throughput per clock as a RDNA4 WGP.

 

[marees]

Kepler_L2 dropped the Transformer names in an AnandTech forum in response to user @igor_kavinski , who noted that the AT0 codename was "boring." Kepler_L2 listed Alpha Trion, Ultra Magnus, and Orion Pax, who eventually became Optimus Prime.

There Might Be More Than Meets the Eye With AMD RDNA 5

AMD appears to be code-naming its next-generation GPUs after Transformers.

www.extremetech.com

 

[marees]

3DCenter has speculated that this method might facilitate the production of GPUs featuring multiple GPU chiplets in the future. However, notable leaker Kepler_L2 has indicated that AMD does not intend to pursue this with RDNA 5.

(Fwiw, it seems to me that AMD has all the puzzle blocks in place with RDNA 5. If the software works fine then it's just getting the hardware to play ball in RDNA 6 or 7 🤔)

Radeon’s upcoming RDNA 5 chiplet approach could reshape AMD’s GPU strategy

Rumors suggest that AMD is planning to reintroduce a chiplet-based approach for its RDNA 5 graphics architecture.

www.dlcompare.com

 

[soresu]

3DCenter has speculated that this method might facilitate the production of GPUs featuring multiple GPU chiplets in the future. However, notable leaker Kepler_L2 has indicated that AMD does not intend to pursue this with RDNA 5.

(Fwiw, it seems to me that AMD has all the puzzle blocks in place with RDNA 5. If the software works fine then it's just getting the hardware to play ball in RDNA 6 or 7 🤔)

Radeon’s upcoming RDNA 5 chiplet approach could reshape AMD’s GPU strategy

Rumors suggest that AMD is planning to reintroduce a chiplet-based approach for its RDNA 5 graphics architecture.

www.dlcompare.com

If the rumors were true then AMD already had them with RDNA4 and simply decided it wasn't the best way to go for that generation.

Given RDNA5 is already going to be a pretty radical change even without SED chiplets it makes sense that they wouldn't be doing that anyway, just like Zen1.

 

[soresu]

"AMD’s RDNA 5 GPUs could be much bigger than expected. According to the leaker zhangzhonghao, AMD has changed the structure of its Compute Units (CU) with RDNA 5. Instead of having 64 Stream Multiprocessors (SM) per Compute Unit, RDNA 5 reportedly features 128. That’s a 2x increase in SM count per CU."

I'm really new to these more indepth topics, this has something to do with the WGP vs CU previous talk? This 2x is really likely to hapen?

I mean, when you actually look at the layout of the RDNA4 WGP it's a bit more than that....

128 FMA/Int dual purpose ALUs + 128 FMA only ALUs for 256 FMA ALUs total, and 32 transcendental logic units (TLUs).

 

[MrMPFR]

128 FMA/Int dual purpose ALUs + 128 FMA only ALUs for 256 FMA ALUs total, and 32 transcendental logic units (TLUs).

If you make that FMA/INT for all ALUs AMD can claim 2x cores for entire lineup like NVIDIA.

Repeating myself (posted a week ago) but maybe that's where 256 cores per CU for GFX12.5 (CDNA5) comes from.


And yes I did change the username.

 

[adroc_thurston]

If you make that FMA/INT for all ALUs AMD can claim 2x cores for entire lineup like NVIDIA

You're not ALU-limited anywhere anyway.

 

[regen1]

https://www.investing.com/news/transcripts/amd-at-goldman-sachs-conference-ai-and-data-center-insights-93CH-4230321

It’s all culminating in our MI450 generation, which we’re launching next year, where that is for us our, you know, no asterisk generation, where we believe we are targeting having leadership performance across the board, any sort of AI workload, be it training or inference. Everything that we’ve been doing has been focused on the hardware and the software, and increasingly now at the system and cluster level as well, to build out that capability so it all intersects. MI450 is perhaps akin to our Milan moment for people that are familiar with our EPYC roadmap.
The third generation of EPYC CPUs is the one where we targeted having no excuses. It was superior. Rome and Naples were very good chips, and they were highly performant and the best possible solution for some workloads. Milan is where it was the best CPU for any x86 workload, period, full stop. We’re trying to view and plan for MI450 to be the same. It will be, we believe, and we are planning for it to be the best training, inference, distributed inference, reinforcement learning solution available on the market.

 

[soresu]

https://www.investing.com/news/transcripts/amd-at-goldman-sachs-conference-ai-and-data-center-insights-93CH-4230321

That's part of a longer interview with the GPU division head.

 

[Joe NYC]

That's part of a longer interview with the GPU division head.

No, this was at Goldman Sachs investor conference, presented by Forrest Norrod.

 

[stayfrosty]

So, what do you guys think AMD will do in response to Rubin CPX?

1. Introduce their own low-precision optimized co-processors based on tweaked client chips like nvidia
2. Introduce MI500(?) flavours that replace part of (or the complete) HBM bus with more compute chiplets and a gddr7 PHY
3. No response at all and give up the high context inference market for the moment

 

[marees]

So, what do you guys think AMD will do in response to Rubin CPX?

1. Introduce their own low-precision optimized co-processors based on tweaked client chips like nvidia
2. Introduce MI500(?) flavours that replace part of (or the complete) HBM bus with more compute chiplets and a gddr7 PHY
3. No response at all and give up the high context inference market for the moment

Will be based on what their clients ask for

Ex: RDNA is driven by Sony

CDNA starting MI400 (next year is already split in two for HPC vs AI)

MI500 is 2027. 2 yrs out

 

[stayfrosty]

Will be based on what their clients ask for

Ex: RDNA is driven by Sony

CDNA starting MI400 (next year is already split in two for HPC vs AI)

MI500 is 2027. 2 yrs out

Clients will want a CPX like solution. It just makes sense for large context inference.

I don't think AMD will have a response ready with MI400.

 

[Win2012R2]

So, what do you guys think AMD will do in response to Rubin CPX?

What exactly makes GDDR7 card (aka CPX solution) so much better for large context inference? This appears to be normal 202 die that will be relatively cheap with relatively large memory, so using it for inference rather than super expensive HBM versions makes sense.

Equivalent response from AMD will be this AT0 chip.

 

[jpiniero]

What exactly makes GDDR7 card (aka CPX solution) so much better for large context inference? This appears to be normal 202 die that will be relatively cheap with relatively large memory, so using it for inference rather than super expensive HBM versions makes sense.

The scalability I guess?

Or it's just stonks.

 

[adroc_thurston]

What exactly makes GDDR7 card (aka CPX solution) so much better for large context inference?

Better? Oh no. Cheaper? yes.

The scalability I guess?

It's cheaper.