AMD RDNA 4 leak confirms RT performance is getting a boost, but it may not be enough for a major hardware overhaul

A respected insider has revealed more details about the changes AMD is planning to make to ray tracing in its next-generation RDNA 4 GPU architecture. While some of the updates look like they significantly improve certain aspects of the algorithms, the leak also suggests we won’t be seeing a comprehensive overhaul of the ray tracing units.

The person disclosing the information is Kepler_L2 on X (through Wccftech), which has a stellar reputation for providing exact predictions about upcoming GPUs, as well as occasionally publishing whitepaper snippets like in this case.

It must be said that the heavily censored “feature description” is not particularly informative, as there is almost no indication of what exactly has been changed. However, the main one that indicates a well escalate in at least one aspect of ray tracing is the “Double Intersect Ray Tracing Engine”.

The name suggests one of two things: AMD’s RT units in RDNA 4 will either be twice as gigantic as those in RDNA 3, or they will simply be able to process twice as many ray-triangle intersection calculations. Ultimately, it doesn’t really matter How gains will be achieved provided that ray tracing performance is significantly improved.

Not that it’s bad in RDNA 3 — quite the opposite, in fact — it’s just that Nvidia’s current RTX 40-series handles this task significantly better than AMD’s GPUs, so a significant improvement is not only expected, but required.

However, when it comes to ray tracing and performance, it is not Just about supporting a multitude of ray and triangle intersection calculations. Traversing the BVH (Bounding Volume Hierarchy) acceleration structures is also time-consuming and places high demands on the GPU caching system.

To that end, AMD is adding some optimizations to improve things in this area, but the lack of information means it’s impossible to say how essential they are. Most telling, though, is the lack of any nod to AMD using an ASIC (application-specific integrated circuit) to speed up BVH transitions.

In RDNA 2 and 3, this is done via compute shaders running on the GPU cores, and if they’re busy doing that, it means they can’t be used for anything else, especially the ray shaders themselves. For the largest Navi GPUs, this isn’t a major issue — since the chips have a huge number of compute units (CUs) to solve the problem — but as you go down the tiers, the lack of dedicated CUs becomes more noticeable.

Of course, AMD could adding them and that’s what’s hidden in the redacted parts of the document, but if I were in the leak business, that’s the first thing I’d show people. That’s because Nvidia’s ray tracing units contain independent ray and triangle intersection engines AND BVH traverse circuits, hence better performance.

Kepler_L2 also believes that the majority of the RT update will be used in the upcoming PlayStation 5 Pro, although this has been mentioned before. Back then, the confusion was caused by the claim that the PS5 Pro GPU would support BVH8 (transitive shaders running through eight children from each node in the BVH).

That’s twice as speedy as RDNA 3, which could be what the “Double Intersect Ray Tracing Engine” is all about. If that’s the case, don’t expect ray tracing in games to automatically be twice as speedy, as that’s not necessarily a performance bottleneck.

I know that many PC gamers still consider ray tracing to be an unnecessary gimmick, and AMD’s initial implementation of the technology was a sensible choice since they weren’t devoting much hardware space to something that would simply reduce performance that much.

However, by the time RDNA 4 graphics cards hit store shelves, it’s likely that both Intel and Nvidia will also have their next GPUs in stock, and in the case of the latter company, it will be the fourth iteration of RTX technology.

Ray tracing is here to stay, so let’s hope AMD still has something special hidden away that will bring its GPUs up to par with Nvidia’s GPUs.