Video explanation
Glowing chip units build pixels as the interface switches between parallel work, memory waits and divergent branches, making throughput visible as coordinated work, delayed data or different execution paths.
Similar work suits parallel execution
Similar calculations across pixels can be spread over many threads; speedup also depends on layout and workload size.
Memory can limit throughput
Compute units need data, and memory waits reduce useful work. Additional concurrent work can sometimes hide latency.
Divergence affects efficiency
Different paths within a warp may require separate execution and reduce efficiency, without stopping the entire GPU.