AMD will primarily focus on ray-tracing performance with their next-gen Radeon RX 7000 GPUs. The RDNA 3 architecture was initially thought to be a proper chiplet design with multiple compute, and cache/controller dies. However, the latest reports from reputed sources more or less confirm that the top-end Navi 31 die will feature a single Graphics Compute Die (GCD) with 12,228 shaders, and several cache/memory controller dies. This isn’t necessarily bad news, but earlier performance estimates were overhyped.
According to @kopite7kimi, the Radeon RX 7900 XT (based on the Navi 31 die) will be “just” 2x faster than the RX 6900 XT in 4K gaming workloads, but a whopping 3.5x faster in ray-traced titles. That’s some serious ray-tracing grunt. It looks like AMD has put some serious R&D into RDNA 3’s ray accelerators. You can almost certainly expect Tier 3 grade ray-tracing capabilities with the 7000 series cards with the BVH traversal, sorting, and ray-triangle/box intersection testing offloaded to the RAs. In RDNA 2, only the ray-triangle/box testing is performed on the RAs. BVH traversal is handled by the shaders.
Additionally the top-end Navi 31 die may just come with a 384-bit bus, a first for the RDNA graphics architecture. Both the earlier generations topped out at 256-bit. A Radeon RX 7900 XT with 384MB of L3 cache and a 384-bit bus interface should offer some serious bandwidth for 4K gaming. Yet, the 2x target seems quite conservative which makes you wonder. Have the vector capabilties have taken a backseat? Either that, or the leaked memory specifications are inaccurate.