Hardware-Aware Performance Characterization of Small Dense Lindblad Propagation for Near-Term Quantum Control

Abstract

Dense Lindblad propagation at the Hilbert-space sizes of near-term transmon control ($d = 3$, $9$, $27$) sits in a regime where cache boundaries, launch overhead, and fixed latency decide which hardware is useful, not peak arithmetic rate. This paper characterizes that regime on two x86 CPUs, two NVIDIA GPUs, and a low-cost FPGA. Each implementation is stated as an algorithm in $d$ with an idealized cost built from separately measured ceilings, and the measured gaps are attributed to specific mechanisms: thread fork/join, the strided read of the propagator, PCI Express (PCIe) transfers, and propagator re-upload. An end-to-end gradient-ascent (GRAPE-style) benchmark against released QuTiP 5.2.3 shows that fast propagation does not imply fast control optimization: the C path wins by up to two orders of magnitude at $d = 3$ but loses at $d = 27$, where the dense LU solve inside the Padé propagator build takes over $90\%$ of the cost. The FPGA contributes a deterministic-latency operating point, $95$ cycles per step at a stress-characterized $108$ MHz with zero cycle spread over $1000$ trials, and its Q1.15 fixed-point drift is measured on the device. The results are a map of operating regimes for the measured grid, with the propagator build, not propagation, as the dominant cost at the largest size.