AMD to Unify Gaming "RDNA" and Data Center "CDNA" into "UDNA": Singular GPU Architecture Similar to NVIDIA's CUDA

[AleksandarK]

According to new information from Tom's Hardware, AMD has announced plans to unify its consumer-focused gaming RDNA and data center CDNA graphics architectures into a single, unified design called "UDNA." The announcement was made by AMD's Jack Huynh, Senior Vice President and General Manager of the Computing and Graphics Business Group, at IFA 2024 in Berlin. The goal of the new UDNA architecture is to provide a single focus point for developers so that each optimized application can run on consumer-grade GPU like Radeon RX 7900XTX as well as high-end data center GPU like Instinct MI300. This will create a unification similar to NVIDIA's CUDA, which enables CUDA-focused developers to run applications on everything ranging from laptops to data centers.

So, part of a big change at AMD is today we have a CDNA architecture for our Instinct data center GPUs and RDNA for the consumer stuff. It's forked. Going forward, we will call it UDNA. There'll be one unified architecture, both Instinct and client [consumer]. We'll unify it so that it will be so much easier for developers versus today, where they have to choose and value is not improving.

According to Jack Huynh, AMD "made some mistakes with the RDNA side; each time we change the memory hierarchy, the subsystem, it has to reset the matrix on the optimizations. I don't want to do that. So, going forward, we're thinking about not just RDNA 5, RDNA 6, RDNA 7, but UDNA 6 and UDNA 7. We plan the next three generations because once we get the optimizations, I don't want to have to change the memory hierarchy, and then we lose a lot of optimizations. So, we're kind of forcing that issue about full forward and backward compatibility. We do that on Xbox today; it's very doable but requires advanced planning. It's a lot more work to do, but that's the direction we're going."

When AMD originally separated CDNA from RDNA, the company wanted to create two separate entities and thought it would be easier to manage. However, it couldn't be further from the truth. Having two separate teams for optimizations is a nightmare both logistically and engineering-wise. Hence, the shift to a monolithic structure of GPU architectures is beneficial to the company in the long term and will ease the development of newer products with both gaming-focused and compute-focused teams at work. This strategy is similar to NVIDIA's CUDA, which has maintained its architecture line in a single lane, with added special accelerators for AI or/or ray tracing, which AMD also plans to do.

View at TechPowerUp Main Site | Source

 

[Zunexxx]

So, they went from GCN to RDNA because they couldn’t do “best of both worlds” and had to optimize the arch separately, and now they are going back again????

 

[TheDeeGee]

Talking about 10 years late.

 

[ncrs]

I hope this means that ROCm (the equivalent of the CUDA software suite) will provide official support for the entire range of next-gen consumer Radeons unlike what they are doing today - supporting just RX 7900 series. Other models might work or not, and if they don't you're on your own.
A major advantage of CUDA is its hardware and software ecosystem.

 

[Fouquin]

Talking about 10 years late.

10 years too late for what? Something they were already doing 13 years ago? GCN was already a combined architecture with a single programming model.

 

[mrnagant]

Wonder what the anticipated timeline is for the first UDNA part. Has this been in the pipeline for a while, so maybe next generation we will have a merge, or is this new so the merge will start happening in 2-5 years?

They will likely just go down a path of making the CUs more flexible in what designs they offer. Like Hopper has 80B transistors to GF4090 76B. Hopper has only 24ROPs compared to the GF4090s 176. If a datacenter GPU need RT or ROP capability, have your design more flexible where you can remove those components and add more of what is wanted. But the design for the components is shared across the entire stack. From datacenters, to embedded, to gaming to a little 2CU iGPU.

 

[Neo_Morpheus]

So the rumors that RDNA 4 is just a small refresh of sort of RDNA3 and RDNA5 being the real new architecture seems more valid.

So in other words, RDNA5 will really be UDNA1.

Also as mentioned, they concentrated on Ryzen first and that bet paid off, now they are concentrating on Radeon/Instinct.

Lets see how it goes, but it does looks promising.

 

[Apocalypsee]

I hope this means that ROCm (the equivalent of the CUDA software suite) will provide official support for the entire range of next-gen consumer Radeons unlike what they are doing today - supporting just RX 7900 series. Other models might work or not, and if they don't you're on your own.
A major advantage of CUDA is its hardware and software ecosystem.

This is what annoys me the most. Unlike nvidia CUDA that works on any modern card, choosing AMD to use ROCm you need to read which product they supported. IIRC on Windows it support high end RX 6800 series as well, and all 7x00 series.

AMD shouldn't make two different arch in the first place, Vega is good for both workloads just need to tweak the software side and arch

 

[Daven]

I love purposely ambiguous Venn Diagrams. So is AMD going for the intersection of just the RDNA and CDNA or are they going for the super small little intersection of RDNA, CDNA and general purpose?

As for not changing the memory architecture and using the Xbox as an example, I wonder if they are going towards a unified system wide memory architecture, where everything shares a single high bandwidth pool similar to data center GPUs and APUs as well as SoCs like Apple's Mx series.

10 years too late for what? Something they were already doing 13 years ago? GCN was already a combined architecture with a single programming model.

Statements like TheDeeGee's usually are made when you don't really know enough about something but want to sound like you do anyway.

 

[darakian]

I hope this lets them really advance on the software front. I think the GCN analogies area little misplaced here since GCN was of the era where AMD had zero cash to spend on iterating the design. GCN was great when it first landed in the 7k series after all.

 

[Vayra86]

I love purposely ambiguous Venn Diagrams. So is AMD going for the intersection of just the RDNA and CDNA or are they going for the super small little intersection of RDNA, CDNA and general purpose?

Yeah that struck me too lol. They might as well have drawn one big circle. Or just not bothered at all and get on with it already, its not like many people really care until there's an actually killer product actually released, I reckon.

Lets see how it goes, but it does looks promising.

Where did you get the impression they're all hands on deck with RDNA / GPU development now as they were with Zen for CPU? There's no Jim Keller nearby, there's only the usual, we've heard it a dozen times before revamp of AMD's abbreviations... Lots of talk about great plans, and little to show for it.

I'm genuinely interested what you are basing your optimism on, is there a source?

 

[Darc Requiem]

So, post VEGA they split into two separate achitectures and now they are shifting back to a unified architecture. Honestly it makes sense. AMD is smaller than Nvidia with more limited resources. So having to optimize software for two architectures made little sense from a manpower standpoint. The question is will the UDNA architechture be able to perform well in both avenues. Whatever the case maybe, I hope AMD can get the GPU aspect of their business together. GPUs desparately need some short of shake up.

 

[Neo_Morpheus]

Where did you get the impression they're all hands on deck with RDNA / GPU development now as they were with Zen for CPU? There's no Jim Keller nearby, there's only the usual, we've heard it a dozen times before revamp of AMD's abbreviations... Lots of talk about great plans, and little to show for it.

I'm genuinely interested what you are basing your optimism on, is there a source?

I will try to find it, maybe it was either Red Gaming or Moore Law is dead, but it was a rumor that RDNA4 is a refresh/error fixing of RDNA3 and hence why we are observing so "little" on AMD part and the reason for all this was the "fact" that RDNA5 was going to be a new design/initiative on their part.

So to me, the last two announcements (abandoning the high end and now this one) kind of align with that rumor.

Edit maybe was this article:

Latest performance rumor around AMD’s RDNA 4 GPUs could worry you – but we think there’s no need to panic

RDNA 5 is what AMD is really focused on – but Team Red will still have to make RDNA 4 GPUs work

www.techradar.com

 

[TumbleGeorge]

So in other words, RDNA5 will really be UDNA1.

What? No sense to name something rDNA if isn't rDNA only. Maybe next generation after rDNA 5.

 

[Neo_Morpheus]

What? No sense to name something rDNA if isn't rDNA only. Maybe next generation after rDNA 5.

Maybe they do end up changing the names or maybe one of those 2 names is just an internal thing, since RDNA is more known, maybe they dont want to abandon it.

 

[z1n0x]

RDNA, CDNA, XDNA, UDNA. This overuse of DNA is so dumb. The sheer lack of taste and imagination to come up with original terms, that are cool, catchy, tells the story and roll of the tongue is astonishing. Then add to it other marketing atrocities like Adrenalin, AFMF, Anti-Lag, HYPR-RX. I swear AMD's marketing is beyond salvation.

 

[Vayra86]

RDNA, CDNA, XDNA, UDNA. This overuse of DNA is so dumb. The sheer lack of taste and imagination to come up with original terms, that are cool, catchy, tells the story and roll of the tongue is astonishing. Then add to it other marketing atrocities like Adrenalin, AFMF, Anti-Lag, HYPR-RX. I swear AMD's marketing is beyond salvation.

If I look at that list I'm thinking 'Wow, those are a lot of designer drugs'!

Advanced MicroDosing?

 

[TumbleGeorge]

Maybe they do end up changing the names or maybe one of those 2 names is just an internal thing, since RDNA is more known, maybe they dont want to abandon it.

It makes sense to have a truly all-new architecture by the time the next generation of consoles is released. Thus, there will be more time for the new architecture to be well baked. Because rDNA itself has more to give on its own that seems to need a more general redesign. It is possible to include in rDNA 5 elements, for the purpose of testing in real conditions, which participate in the subsequent integration with cDNA.

 

[renz496]

sometimes you just want to say like "AMD please make up you mind already". they saw what nvidia did with kepler/maxwell and pascal/volta a few years ago and they try to do something similar to that. nvidia go back with compute architecture with turing while AMD decided to go RDNA/CDNA. now they go back with "unified" architecture strategy where pro and consumer share the same architecture just with their own tweak like nvidia currently did.

thing is nvidia kepler/maxwell and pascal/volta works for them because nvidia still properly support CUDA on their card regardless the card's architecture focus. with AMD when they go RDNA/CDNA route we saw that AMD try to support ROCm mostly on CDNA only. i still remember polaris and vega have some sort of ROCm support but effort is obviously lacking on RDNA and RDNA2.

 

[Makaveli]

So the rumors that RDNA 4 is just a small refresh of sort of RDNA3 and RDNA5 being the real new architecture seems more valid.

So in other words, RDNA5 will really be UDNA1.

Also as mentioned, they concentrated on Ryzen first and that bet paid off, now they are concentrating on Radeon/Instinct.

Lets see how it goes, but it does looks promising.

Based on what he said I think it will start at RDNA 6 not 5.

 

[Neo_Morpheus]

Based on what he said I think it will start at RDNA 6 not 5.

Its possible.

From Toms:

JH: So, one of the things we want to do is ...we made some mistakes with the RDNA side; each time we change the memory hierarchy, the subsystem, it has to reset the matrix on the optimizations. I don't want to do that.

So, going forward, we’re thinking about not just RDNA 5, RDNA 6, RDNA 7, but UDNA 6 and UDNA 7. We plan the next three generations because once we get the optimizations, I don't want to have to change the memory hierarchy, and then we lose a lot of optimizations. So, we're kind of forcing that issue about full forward and backward compatibility. We do that on Xbox today; it’s very doable but requires advanced planning. It’s a lot more work to do, but that’s the direction we’re going.

Interesting how he mentions RDNA5 but then says UDNA6, so it might be the case.

But if thats the case, then RDNA5 will simply be more of the same.

 

[Makaveli]

Its possible.

From Toms:

Interesting how he mentions RDNA5 but then says UDNA6, so it might be the case.

But if thats the case, then RDNA5 will simply be more of the same.

I'm thinking RDNA 5 may already be too far in development but just a wild guess on my part.

 

[Flyordie]

So, post VEGA they split into two separate achitectures and now they are shifting back to a unified architecture. Honestly it makes sense. AMD is smaller than Nvidia with more limited resources. So having to optimize software for two architectures made little sense from a manpower standpoint. The question is will the UDNA architechture be able to perform well in both avenues. Whatever the case maybe, I hope AMD can get the GPU aspect of their business together. GPUs desparately need some short of shake up.

This is why Vega was still getting driver updates this far out. Vega's compute units are the same as the ones used on current Instinct cards. All CDNA is are the graphics engine components ripped out of GCN5.x and the associated tweaks made to optimize for that new die layout.

 

[Minus Infinity]

Wonder what the anticipated timeline is for the first UDNA part. Has this been in the pipeline for a while, so maybe next generation we will have a merge, or is this new so the merge will start happening in 2-5 years?

They will likely just go down a path of making the CUs more flexible in what designs they offer. Like Hopper has 80B transistors to GF4090 76B. Hopper has only 24ROPs compared to the GF4090s 176. If a datacenter GPU need RT or ROP capability, have your design more flexible where you can remove those components and add more of what is wanted. But the design for the components is shared across the entire stack. From datacenters, to embedded, to gaming to a little 2CU iGPU.

He gave you a clue. UDNA6. RDNA 5 is already locked in and being developed and progressing well. So at the very earliest it would be like 2027 when one would expect RDNA6 normally. But it might be the gen after, so 2028/29.

I'm thinking RDNA 5 may already be too far in development but just a wild guess on my part.

Not a wild guess, RDNA5 is very advanced, not a snoball's chance in hell they'd abandon it at this stage. It takes at least 3 years to design new gpu and that's starting with an existing previous design.

 

[1d10t]

So AMD went back to pre-RDNA or GCN/Polaris era, personally I don't see that as a win nor a white flag, but I do hope they understand needs of market segmentation, not just plain cut down. As we already knew Vega, which I used two of them in Crossfire, is a bit underwhelming with its power consumption, HBM only serve as bragging and doesn't really make a big differences. Hopefully this time around they know what they are doing.