AMD’s Radeon RX 7900 XT is likely going to be 40-80% faster than its predecessor, as per Moore’s Law is Dead in a new video. This falls in line with what I said in my last post on the RDNA 3 flagship. Furthermore, there’s a chance that the next-gen graphics cards from AMD may not feature a truly chiplet design. Although the GPUs are going to feature at least two dies, one of them is going to be the I/O die, handling the various output ports, PCIe lanes, etc. My guess is that the flagship will feature two compute dies and one I/O die, while all the other SKUs will consist of just one compute and I/O each.
This means that if Navi 33 is a dual-die design, then it should be 15-20% faster than the RX 6900 XT (Navi 21), while Navi 31 (RX 7900 XT) will be 50-80% faster thanks to two compute dies and one I/O die. Considering that GPUs are much more sensitive to latency than CPUs, the scaling is not going to be linear, indicating that we will see a performance gain of less than 2x on the top-end.
Ray-tracing performance is expected to improve with RDNA 3 with an increased amount of die space dedicated to acceleration/intersection units, along with geometry performance which saw a drop with RDNA 2. Each RDNA 2 Shader Engine packed just a single primitive unit, less than half as much on NVIDIA’s Ampere GPCs, and the original RDNA 1 design. This was primarily done in anticipation of enhanced culling brought about by mesh shading, one of the primary features of DX12.