AMD's software engineers have been very busy lately to update linux GPU kernel drivers, shader compilers, and other code sources to support the next generation of graphics architectures. Now the same thing has been done with the RadeonSI OpenGL driver, with almost 12 lines of code for GFX4 (aka RDNA4) added to the Mesa open source graphics library.

Large code merges were discovered by Phoronix, but the reason engineers are busy updating the codebase for architectures that are not currently on the market is because

The current GPU chip design, RDNA3, is codenamed internally by AMD as gfx11. The reference to GFX12 is clearly its successor I'm not talking about RDNA11.5, which is listed as GFX3.5, but only brings minor changes when it appears on AMD's Strix Point and Strix Halo laptop Apu. All the recent code merges are for RDNA4, but the file itself doesn't tell me much about what you can expect from the architecture going forward. The Radeon RX7000 series Gpus will be first introduced in 2022-11, while the RX7900XTX andXT are the first consumer-level graphics cards to use chiplets. Despite having a lot of performance, AMD is struggling to sell as many RDNA2 cards as it does RDNA3 cards.

Things are a bit rosy thanks to reduced inventory for certain RX6000 models and various price cuts such as the RX7900XT and 7800XT, but overall, AMD is in the discrete GPU market, what can or can be expected of RDNA4 — or another way to ask this is to improve sales. What we need to improve on AMD's next GPU architecture is what we need to do to improve the performance of the GPU. There's no problem with the basic rendering performance of AMD's design (aka rasterization), which makes RDNA3 the area behind Nvidia's Ada Lovelace-powered RTX40 series. This is ray tracing, which leverages machine learning to improve performance and power efficiency. And that's where rumors are pegging AMD's new GPU architecture as something different from the current generation; Another way to approach ray tracing in RDNA4. 

RDNA3Gpu has dedicated hardware for accelerating ray-triangle intersections, but nothing for doing BVH traversal - it's all done through the same unit that handles everything in a normal shader. This is a similar situation for performing deep learning neural network calculations. If Intel and Nvidia have large dedicated matrix units, AMD will use a combination of "AI accelerators" and shader units to achieve the same. Things like the Radeon RX7900XTX have the same average power consumption as Nvidia's equivalent cards, but peak power demand is usually much higher, similar to idle consumption. Multiple chiplets and complex infinity link systems are ideal for wafer yields, but their initial implementation all require fine tuning to improve power efficiency.

However, there is an increasing possibility that very high-end RDNA4 graphics cards will not appear or will not appear at launch. There was a lot of talk about AMD skipping the halo sector for the next GPU release and focusing more on the midrange and mainstream markets.This refers to sticking to a monolithic design or one with a slight chiplet.

Whatever changes AMD has prepared for us in RDNA4, everything is set, at least from a driver perspective. 

And best of all, it doesn't seem to take much time to wait to find out how good it is.