CUDA Bottlenecks
Medium · Phone screen · Onsite · Full-time
They asked how I'd go about identifying performance bottlenecks in a CUDA kernel. I talked about using Nsight Compute to look at occupancy, memory throughput, and instruction throughput. The interviewer was particularly interested in memory coalescing, whether I understood shared memory bank conflicts, and how I'd decide if a kernel is memory-bound or compute-bound. We also briefly discussed profiling workflow and what metrics I'd check first.
