Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data | Hacker News

GPU matrix multiplication can run faster when the input data is more predictable, suggesting workload patterns affect performance beyond raw compute speed.

The discussion focuses on why this happens, including speculation about prediction, execution efficiency, and how GPUs handle control flow.

Commenters also raise concerns about wasted power at idle and the broader efficiency cost of keeping GPUs ready for heavy workloads.

They argue that power caps or memory limits could force better software and hardware optimization, with implications for AI, gaming, and other compute-intensive uses.