Your question is Profile GPU Underutilization Signals. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How would you profile a CUDA application to identify GPU underutilization, and which timeline and utilization metrics would you inspect first to determine whether the bottleneck is on the CPU, launch path, memory transfers, or kernel execution?
