How to Improve CUDA Kernel Performance with Shared Memory Register Spilling

How to Improve CUDA Kernel Performance with Shared Memory Register Spilling

How to Improve CUDA Kernel Performance with Shared Memory Register Spilling | NVIDIA Technical Blog

“When a CUDA kernel requires more hardware registers than are available, the compiler is forced to move the excess variables into local memory, a process known as register spilling. Register spilling affects performance because the kernel must access local memory—physically located in global memory—to read and write the spilled data.

In CUDA Toolkit 13.0, NVIDIA introduced a new optimization feature in the compilation flow: shared memory register spilling for CUDA kernels. This post explains the new feature, highlights the motivation behind its addition, and details how it can be enabled. It also provides guidance on when to consider using it and how to evaluate its potential impact…”

Source: developer.nvidia.com/blog/how-to-improve-cuda-kernel-performance-with-shared-memory-register-spilling