Discussion RDNA 5 + CDNA 4 Architectures Thread

[Kronos1996]

Now seems like a good time to start this discussion thread. Per AMD’s earnings call yesterday, CDNA 4 is coming out mid-2025. MI350 sounds like it uses the new architecture with the same packaging/platform while MI400 is a top to bottom redesign.

RDNA 5 seems to be 18 months away minimum (late 2026 launch per leaked roadmaps.) Still, they should have the architecture locked in. Might get some info on the dies later this year when they start tape-out’s.

My main question is if these two architectures are the first step in the UDNA merge? Same base architecture heavily customized for gaming and server. Essentially what Nvidia’s been doing for years.

Not that dissimilar to AMD’s approach with Zen either. Classic core, Dense Core and modular add-ons like X3D. They can add/remove IP like AVX-512 depending on the market. Most importantly, all the variants are software compatible.

It sounds like they want to take a similar approach with UDNA. Unified software stack for a unified base architecture. My guess is, a new modular CU. Then they can insert whatever IP is appropriate for each product line. Rasterization + RT hardware for gaming. The whole shopping list of AI instructions for datacenter. Optimized density for each depending on the clock-speed targets.

David Wang made remarks about how changing the cache system resets software optimization to zero. Is it safe to assume both products lines will share a similar memory system design?

 

[marees]

There is a one / one-and-a-half year gap between CDNA 4 & RDNA 5

So maybe the one after CDNA 4 could be the UDNA version ??

 

[Kronos1996]

There is a one / one-and-a-half year gap between CDNA 4 & RDNA 5

So maybe the one after CDNA 4 could be the UDNA version ??

Could be yeah. I thought MI350 was a relatively new addition and they originally planned CDNA 4 to debut with MI400? If that’s the case, then CDNA 4’s launch has been moved up considerably.

 

[SolidQ]

There is already topic, don't need duplicate

 

[marees]

There is already topic, don't need duplicate

Apologize for the confusion.

I see that there is no dedicated thread for CDNA 4. Maybe this thread can be renamed CDNA 4 only (& remove RDNA 5?)

AMD Plans Massive Memory Instinct MI325X for Q4'24, Lays Out Accelerator Roadmap to 2026

www.anandtech.com

 

[GTracing]

My main question is if these two architectures are the first step in the UDNA merge?

There is a one / one-and-a-half year gap between CDNA 4 & RDNA 5

So maybe the one after CDNA 4 could be the UDNA version ??

Jack Huynh hinted that the RDNA5/CDNA4 generation won't be fully unified when he announced the UDNA name. It sounds like it will called UDNA regardless of whether RDNA and CDNA are merged into one architecture or not.

PA: So, this merging back together, how long will that take? How many more product generations before we see that?

JH: We haven’t disclosed that yet. It’s a strategy. Strategy is very important to me. I think it’s the right strategy. We’ve got to make sure we’re doing the right thing.

AMD announces unified UDNA GPU architecture — bringing RDNA and CDNA together to take on Nvidia's CUDA ecosystem

Two become one.

www.tomshardware.com

 

[basix]

MI400 should already be CDNA5. So maybe RDNA6 & CDNA5 = UDNA? Huynh mentioned framed it as a cloud-to-client strategy, which wold match that. CDNA5 is first and RDNA6 follows. RDNA5 could be still somewhat close to RDNA2 which might be better for next-gen consoles also.

 

[CakeMonster]

If this is the basis for hardware in the next gen consoles I really hope they can cram as much 'universal' hardware units as possible in there so that it will run future stuff we can come up with as AI models are trained in 1-2-5 years, similar to transformer based upscaling now running on 2018 Turing hardware.

 

[rommelrommel]

If this is the basis for hardware in the next gen consoles I really hope they can cram as much 'universal' hardware units as possible in there so that it will run future stuff we can come up with as AI models are trained in 1-2-5 years, similar to transformer based upscaling now running on 2018 Turing hardware.

The transformer model runs on Turing, but so poorly you won't use it.

 

[marees]

CDNA 4 will deliver the biggest generational leap in AI performance in our history, with a 35 times increase in AI compute performance compared to CDNA 3.

Advanced Micro Devices (AMD) Q4 2024 Earnings Call Transcript | The Motley Fool

AMD earnings call for the period ending December 31, 2024.

www.fool.com

 

[Thunder 57]

The transformer model runs on Turing, but so poorly you won't use it.

Just saw the HUB video on this. Even a 2080 Ti is basically unusable.

 

[CakeMonster]

Just saw the HUB video on this. Even a 2080 Ti is basically unusable.

Ray Reconstruction and Upscaling are different, the Transformer model works fine for the latter.

 

[soresu]

CDNA 4 will deliver the biggest generational leap in AI performance in our history, with a 35 times increase in AI compute performance compared to CDNA 3.

Lisa Su in the CDNA4 product development meetings.....

 

[marees]

CDNA4 also brings native support for FP6 and FP4 data types to AMD’s accelerators for the first time. One of the marquee features of rival NVIDIA’s Blackwell architecture, FP6 and FP4 have become a new target for AI inference, as developers look to wring every TOP/FLOP of performance from these expensive and power-hungry GPUs. And, aiming to one-up NVIDIA at their own game here, AMD has even beefed up FP6 performance on their architecture so that it processes at twice the rate of FP8, unlike NVIDIA’s architecture where it processes at the same rate as FP8. AMD in essence built a better FP4 unit to support FP6, rather than reusing an FP8 unit to support FP6. This carries a die area penalty, but the upshot is double the performance.

https://twitter.com/x/status/1933233181039472709

AMD's AI Future is Rack Scale 'Helios'

Key Announcements from AMD Advancing AI 2025

morethanmoore.substack.com