What happens when a GPU reads memory
Summary
The article provides a deep dive into the GPU memory access path on an RTX 4090, tracing a global memory load from warp execution through L1, L2, translation lookaside buffers, and DRAM. It includes latency figures (e.g., L1 ~15.4 ns, L2 ~127 ns, DRAM ~255 ns) and a detailed narrative of how data travels through the memory hierarchy, along with an appendix of probes and measurements. This is valuable for understanding hardware-level performance characteristics and memory bottlenecks in GPU kernels.