RTX 5090 Sits Idle at 23% While CPU Solves Go Puzzles, Pointing to L2 Cache Gains

A benchmark of the RZS-TT Go solver — a program that exhaustively proves life-and-death outcomes rather than playing moves — found the RTX 5090 GPU nearly idle at around 23% utilization throughout an 11-hour run. The solver's hot path runs entirely on the CPU, performing random, pointer-chasing lookups in a transposition table whose speed depends on cache latency, not GPU compute. Running on a Core Ultra 9 285K, the solver proved 84 of 117 puzzles (72%) within a five-minute-per-problem budget, compared to 68 of 106 (64%) on the paper's reference Xeon E5-2683 v3 from roughly a decade ago. The most striking hardware difference between the two machines is private L2 cache per core, which grew approximately 12-fold from 256 KB on the older Haswell chip to 3 MB on Arrow Lake, while shared L3 remained nearly flat at around 35–36 MB. The author argues this large per-core L2 expansion likely kept more of the hot transposition table closer to each thread, reducing costly cache spills — though the piece acknowledges this remains a hypothesis rather than a controlled measurement.