AMD RDNA 5 GPUs feature Redesigned Compute units and Advanced Cache for High Performance

AMD’s next-generation GPUs integrate shared L0-LDS cache and high-bandwidth memory controllers to support complex shader workloads and AI processing.

Hardware by Tanvir Kabbo on Sep 26, 2025

Next generation of AMD graphics cards is shaping up to be very, very interesting indeed. Interestingly, they are taking a more Nvidia-like approach. This isn't just in terms of the cache structure, but also in how they are organizing the compute units and workgroup processes.

As we get deeper into the details, you'll see what I mean. The information comes courtesy of Kepler L2, who posted on the Anandtech forums. Throughout this discussion, I'll be referring to these cards as RDNA 5 or UDNA, though no one is 100% certain what they will ultimately be called.

RDNA 5 is the closer technical name, but the GFX13 series will have multiple iterations, including modifications for AI, server applications, and other specialized tasks. This will likely create some confusion between the various iterations.

A few days ago, rumors surfaced regarding the 83 and 84 models, which represent mid-range and lower-end specifications. You'll get more context on this shortly.

AMD, RDNA 5 GPUs, feature Redesigned Compute units, Advanced Cache, High Performance, NoobFeed

Changes in Compute Units and Workgroup Processors

Focusing on the flagship 80 model, the compute units immediately catch your attention. At first glance, counting from the top left, you see 1, 2, 3, 4, 5, 6, and counting across all sections gives a total of 96 compute units. The big question is: why?

According to rumors, the traditional workgroup processes are being removed. Instead of having a parent-like structure where compute units nest under a workgroup processor, each compute unit now acts as the central authority, with shaders organized beneath it.

This change will have intriguing ramifications for cache usage and shader access to data within those caches. In this configuration, there are six shaders per shader array for the AT structure, giving eight shader arrays with two shader engines under each. This results in a total of 16 shader engines, each containing six compute units.

Memory and Cache Configuration

The maximum rumored configuration includes a 512-bit memory interface and allegedly 128MB of Infinity Cache. However, this remains unconfirmed, with rumors suggesting the cache size may either shrink significantly or remain similar to previous iterations. These changes will be especially interesting as shader complexity increases.

Kepler mentions that the L0 and LDS (Local Data Store) cache setup from CDNA4, which utilized 32KB L0 and 160KB LDS, could be expanded to a shared 448KB L0-LDS cache in CDNA5. This merging of caches appears designed to effectively handle the increased number of shaders per compute unit.

AMD, RDNA 5 GPUs, feature Redesigned Compute units, Advanced Cache, High Performance, NoobFeed

Variations Across Models

The 82 model is slightly smaller, featuring a total of 40 compute units, five compute units per shader engine, and four shader arrays. With six memory controllers, this gives a theoretical 192-bit memory interface, though exact configurations remain uncertain.

The 83 model decreases further to 24 compute units, while the 84 is the smallest in the lineup. Importantly, each compute unit in the 80 model is roughly equivalent to two compute units from the previous generation, meaning the 96 compute units effectively perform like 192 from the earlier architecture.

Rumors and Source Reliability

Zang Xong Hua from the Chip Health forums agrees with the titles "extra large," "medium," "small," and "tiny," which contribute to some of these aspects. There are a lot of speculations going around concerning RDNA5, though, and the information may not be credible because it comes from so many different places.

There is also speculation that the highest-end configurations of AMD's next-generation GPUs may be exclusive to data centers, rather than being available to general consumers. Until official details emerge, performance projections remain speculative, as factors like clock speeds and exact memory layouts are unknown.

Competition and Industry Outlook

Nvidia's response with the RTX 60 series will be crucial. While certain improvements over RTX 50 are expected, such as enhanced ray tracing performance and AI-focused tensor cores, space limitations on the die will always impose constraints.

Efficient bandwidth, memory controllers, and cache management will all influence performance.

For AMD, RDNA 5 theoretically has the potential to outperform an RTX 5090, especially with its 96 compute units. However, the outcome depends on how well the cores are fed, how bandwidth contention is handled, and performance in ray tracing scenarios.

Numerous patents, some of which have emerged recently, suggest that AMD is actively developing ray tracing optimizations, although many details remain under wraps.

AMD, RDNA 5 GPUs, feature Redesigned Compute units, Advanced Cache, High Performance, NoobFeed

Final Thoughts

The next generation of AMD GPUs is likely to be quite competitive. Changes to the way compute units, shader arrays, and caches are set up could have a big effect on performance.

There are still certain things that aren't clear about memory setup, clock speeds, and Nvidia's response, but both cons and pros should keep a watch on RDNA 5 as more official information comes out.

Check Our Other AMD articles: