RDNA 4's “Out-of-Order” Memory Accesses

AMD's RDNA 4 architecture introduces significant enhancements to its memory subsystem, particularly in handling out-of-order memory accesses. This new capability allows requests from different shader waves to be processed independently, addressing a limitation in RDNA 3 where memory requests were handled in a strict order, leading to false dependencies between waves. The testing conducted revealed that RDNA 4 effectively eliminates these cross-wave delays, allowing for improved performance, especially in workloads like ray tracing. The architecture also features multiple out-of-order queues for memory requests, enhancing the efficiency of memory access handling within a wave. This change allows threads to interleave different types of memory requests, improving overall throughput. While RDNA 4's memory management improvements are notable, they are seen as evolutionary rather than revolutionary, as similar techniques have been implemented in other GPU architectures from Intel and Nvidia. Overall, RDNA 4 represents a significant step forward for AMD's GPU memory subsystem, enhancing performance across various applications.

- RDNA 4 allows out-of-order memory accesses, improving performance by eliminating false dependencies.

- The architecture introduces multiple out-of-order queues for memory requests, enhancing efficiency.

- Improvements are particularly beneficial for ray tracing workloads, allowing simultaneous traversal and result handling.

- While significant, RDNA 4's enhancements are seen as evolutionary, with similar features present in other GPU architectures.

- The changes mark the most substantial update to AMD's GPU memory subsystem since the launch of RDNA in 2019.