tl;dr: When a CPU appears 100% utilized, it's often stalled waiting on DRAM (50-100ns latency vs ~1ns for L1 cache). The author demonstrates this with a GROUP BY query where perf shows 50% of CPU slots stalled on memory, then applies radix partitioning to split the hash table into 4096 smaller tables that fit in cache—cutting DRAM reads from 4.5GB to 1.5GB and runtime from 3.6s to 2.7s. Their Pivot engine, with further optimizations, hits 1.8s single-core and 171ms across 16 cores.