The Same RTX 5090, but the GPU Sat Idle — a CPU-Bound Go Solver and the Case for L2 Cache

Short summary

A benchmark of a Go life-and-death solver on an RTX 5090 reveals the GPU sits idle at ~23% utilization while CPU cache hierarchy is the real bottleneck. Comparing a 2024 Core Ultra 9 285K against a decade-old Xeon, the 12× growth in per-core L2 cache (256KB to 3MB) is the standout factor explaining improved solve rates (72% vs 64%). The article argues L2 cache size, not GPU power, drives performance for pointer-chasing transposition-table workloads.

• •RTX 5090 sits idle at ~23% utilization in a CPU-bound Go solver benchmark
• •Per-core L2 cache grew 12× (256KB→3MB) over a decade, likely explaining improved solve rates
• •Shared L3 cache remained essentially flat (35→36MB), pointing to L2 as the key differentiator

Generated with AI, which can make mistakes.