AMD RDNA5 Deep Dive Architecture Changes Consoles and Future GPUs

RDNA5 development reveals deep architectural changes driven by shared cache design and next generation rendering demands.

Hardware by Shinji Okazaki on Jan 26, 2026

AMD's GFX13, also known as RDNA5, is going to change how graphics chips are built going forward. Expectations are that ray tracing, rasterization, and power efficiency will all get better. This architecture is likely to serve as the basis for AMD's next-generation discrete GPUs and for future consoles, namely the PlayStation 6 in both handheld and home versions, and the next-generation Xbox.

Console makers are likely to make their own changes, but the basic design already looks rather solid. Leaks and patent activity are starting to give us a clearer idea of where AMD is going with its next-generation GPUs. Early signs point to high-end chips being able to compete with Nvidia's forthcoming offerings.

AMD, RDNA5 Deep Dive Architecture, Changes Consoles and Future GPUs, NoobFeed

Early Mentions of GFX13 and New LLVM Features

AMD has added GFX13 references to LLVM, which means that Linux support is coming soon. Entries like GFX1310 are for future discrete GPU configurations that might be related to designs in the 80 or 82 class. People had issues about the compute unit mode; research from many sources indicated that this might be a placeholder or a copy-paste error.

More modifications mean that RDNA5 works more like MI400, with no dynamic switching between computing units and workgroup processors. This makes both units functionally the same. This makes the execution model easier to understand and fits with the idea that workgroup processors and computing units are the same thing.

Count and Divide up Compute Units

Leaked information and mockup diagrams suggest an 80-class RDNA5 GPU with 96 compute units or workgroup processors. Some estimates claim higher numbers, such as 154, though these differences are likely due to segmentation during early design stages. No matter how many there are, it's evident that AMD is looking into a wide range of setups. These uncertainties also affect the memory subsystem, where further changes are likely.

GFX13 Variants and Console Goals

More discussion shows that GFX1350 is CDNA6, which is mostly important for computing and data center goods. GFX1300 and GFX1301 are two more consumer-friendly versions, thought to be aimed at Orion and Canis designs. These go with the PlayStation6 home console and portable systems. Recent GitHub activity also shows instruction-level enhancements, such as SOP and VOP1 operations, indicating that the ISA is still growing. It's not clear what the full feature set will be, but these new features show that development is still ongoing.

What We Learned from the PlayStation5 Pro Architecture

PlayStation 5 Pro technical session provided background on AMD's evolving approach to GPU architecture. That GPU is a mix of RDNA2, RDNA3, and RDNA4, making it a hybrid design well-suited to console workloads. Vector registers were combined to provide neural network workloads with sufficient bandwidth. This resulted in total memory and bandwidth of about 15 MB and 200 TB/s, respectively. These choices show that AMD is more interested in making data travel quickly than in just building up typical caches.

Changes to the Cache Hierarchy in RDNA4 and Later: When RDNA4 came out, it got rid of the old L1 cache behavior, which was a big change to the cache hierarchy. L1 is not a separate cache level; it is a file that can both read and write. This method cuts down on unnecessary calls and changes the way data moves through the GPU. Several sources say that RDNA5 has a small L2 cache, so these changes are necessary. Some people think that Infinity Cache might be there, but most evidence shows that it's smaller or not there at all.

Ideas for Patent Analysis and Shared Cache

AMD's most recent patents describe shared-cache architectures intended to improve effective bandwidth and reduce data duplication. These patents describe ways to minimize contention in L2 caches and improve bandwidth efficiency by enabling cores to share cache lines more effectively. The patents emphasize removing unnecessary data and enabling local caches to communicate with each other with as little extra work as possible. The simulation findings in these papers show that shared cache architectures can greatly enhance performance.

Historical Parallels and Simulation Parameters

One patent simulation shows a GPU having 28 computing units, 48KB scratchpads, 16KB L1 caches per unit, and a 1MB L2 cache paired with 8GB of GDDR5. The GPU runs at 1400MHz. These specs are more like early GCN devices like Tahiti or Hawaii than they are like contemporary GPUs. These are just simulation parameters, but they show that AMD is returning to some old ideas to inform future designs, especially regarding cache size and data storage.

Direction for Neural Processing and Rendering

Recently, AMD has made a lot of news about neural arrays, radiance cores, and universal compression. Things like these point to a move toward neural graphics and machine learning acceleration that are more like what Nvidia does. These are not just ideas, as made clear in official speeches; they are part of the planned architecture. For these features to work, the system needs to be deeply integrated. This is why redesigning the cache and bandwidth for RDNA5 is so important.

AMD, RDNA5 Deep Dive Architecture, Changes Consoles and Future GPUs, NoobFeed

The Bigger Picture of the Industry

Changes to DirectX and new rendering techniques further support the hypothesis that ray tracing and neural rendering will become increasingly important in the future. Compared to today's AMD GPUs, ray tracing performance is projected to improve significantly. Aside from GPU improvements, NVIDIA's next-generation RTX 60 architecture is said to bring about big changes that will make the competition much harder. New CPUs like the 9950X3D 2 and rumors about Zen6 also point to a longer transition period for PC hardware.

Final Thoughts

There's more to RDNA5 than just a small change. There needs to be a big rethink because the architecture of computing units, the hierarchy of caches, shared memory, and neural processing have all changed. Patents, leaks, and new software support all point to AMD getting ready for a big change in GPU design that will be good for games, consoles, and computers in the future.

Check Our Other AMD articles below: