Pivotlake explains why 100% CPU utilization does not mean useful computation fills every cycle. In a single-core AMD EPYC 9124 experiment grouping roughly 99 million rows into 17 million groups, 4,096-way radix partitioning improved locality and reduced warm execution time from 3.6 to 2.7 seconds. Instruction count increased, but the memory-stall metric fell from 50.1% to 23.5%. DuckDB likewise uses small local hash tables and partition-wise aggregation to reduce cache misses and memory pressure. These are workload- and hardware-specific results, supporting measurement of the actual bottleneck before adding cores.