What are the differences among Compute (SM) Throughput, Memory Throughput, and DRAM Throughput in GPU Speed Of Light Throughput in nsight compute?

The key question is what the source and destination are for each throughput metric? What is the Compute Memory Pipeline Throughput?

Reference:

  1. Compute (SM) Throughput:sm__throughput,SM throughput assuming ideal load balancing across SMSPs (This throughput metric represents the percent of the peak sustained rate achieved during elapsed cycles across all unit instances)
  2. Memory Throughput:gpu_compute_memory__throughputCompute Memory Pipeline Throughput (This throughput metric represents the percent of the peak sustained rate achieved during elapsed cycles across all unit instances)
  3. DRAM Throughput:gpu_dram__throughput,GPU DRAM throughput (This throughput metric represents the percent of the peak sustained rate achieved during elapsed cycles across all unit instances)
  4. DRAM throughput metrics