As you probably already know, AMD recently launched its first two video cards based on the RDNA 3 architecture, namely the Radeon RX 7900 XT and Radeon RX 7900 XTX. Both are equipped with the Navi 31 GPU and offer very good rasterization performance, even superior, as regards the RX 7900 XTX model, to the NVIDIA GeForce RTX 4080. Great strides have also been made in the Ray Tracing field, although it has not yet reached the level of its direct "green" rival. For all the details of the case, we refer you to our full review.
However, according to leaker Kepler, it seems that the first revision of the Radeon RX 7900 cards does not offer truly complete chips and this could have also led to performance lower than expected. In fact, he discovered that early RDNA 3 chips, which include the GFX1100 (Navi 31), GFX1102 (Navi 32), and GFX1103 (APU for the Phenix range), have broken shader prefetch hardware. According to the latest update on GitHub, it seems that the Navi 32 GPUs are actually based on the IP “GFX1103”, while the Navi 33 ones have the IP “GFX 1102”. The problem is also seen on all other chips besides Navi 32 and will certainly also be present in various GPUs that will arrive on both desktops and notebooks early next year.
So AMD decided to release Navi31 A0 silicon, which is known to have a non-working shader prefetch HW ♂️♂️♂️♂️ https://t.co/cQTRKynN0B pic.twitter.com/CIdv6GUcYW
— Kepler (@Kepler_L2) December 14, 2022
According to Kepler, this is not a situation that can be solved in a few months and AMD could take up to a year before launching a new, more efficient revision of these chips on the market. Also, important features like the VOPD instructions featured in RDNA 3 GPUs were supposed to offer a big performance boost, but in reality only managed to deliver a 4% improvement over RDNA 2 in ray traced titles. To this proposition, an AMD representative told Hardware Times colleagues :
Wave64 can natively access the new ALUs for twice the execution speed to unlock performance while executing code Intense ALU. For Wave32 mode, the compiler performs localized reordering and instruction packing in VOPD encoding. An RT test scene using VOPD encoding provided about a 4% increase in frames per second, eliminating the ALU bottleneck. We expect to see more improvements as the compiler matures with more optimizations for mapping code sequences into VOPD encodings. And with advances in using artificial intelligence, RT, and computation-driven rendering techniques for more realistic rendering, we expect to see code that will increasingly take advantage of these new ALUs.
It's possible that AMD was not completely ready to launch its new cards and marketed them so as not to lose too much ground compared to NVIDIA, already present on the market with its GeForce RTX 4090 and RTX 4080.