CPU vs GPU asymmetry — a measured study

Measured constants for an i5-10300H against a GTX 1650 (a 16.4 µs launch round trip, 240 GFLOP/s of thermally capped CPU compute, 32 GB/s sequential and 4 GB/s scattered memory bandwidth) and a crossover at a 64×64 matrix multiply. The reusable piece compiles and runs real CUDA kernels through ctypes with no CUDA toolkit installed.

PeriodSep 2026StatusResearchAreaTools, agents & research

Highlights

Crossover and asymmetry thresholds written up with the caveat that the same algorithm on the GPU wins back 2.4×: the win is algorithm against algorithm.
NVRTC from a pip wheel plus the driver DLL, so GPU experiments on a locked-down laptop need no SDK.
Practical CPU findings: threshold masking over argpartition (7×) and contiguous slices over fancy-index gathers (7.5×).