A Sky-validated RDMA-fed GPU FX Correlator for QUEST

Abstract

We present a reconfigurable, real-time GPU FX correlator for the four-antenna commissioning configuration of QUEST (Qilu University Explorer Survey Telescope). Deterministic RFSoC/FPGA acquisition is paired with a software-defined Python/CuPy correlator on a commodity GPU server: a custom InfiniBand RDMA receiver delivers the 8-bit I plus 8-bit Q, dual-polarization, 250 MHz baseband streams from four antennas directly into pre-registered, page-locked host buffers, and a single NVIDIA RTX 4090 performs PFB channelization, cross-correlation, and integration, with the FFT length, PFB tap count and window, and integration length selectable at runtime. The optimized live path sustains the 32 Gb s−1 active-acquisition line rate, with a 31.05 Gb s−1 wall-clock mean over the full run including startup and shutdown. During a 2 hr Crab Nebula tracking observation on 2026 June 27, the backend processed 29.5 TB of baseband data with no monitored drops in the RDMA-to-F/X path and no F/X evictions. End-to-end phase coherence is verified on sky by comparing the raw, uncalibrated visibility fringes against the geometric delay model: the fringe is detected on all six baselines, and on the 55.3 m N3–N6 baseline the measured group-delay track matches the model with a Pearson correlation coefficient r = 0.9999 (the normalized covariance of the two tracks) and a 0.23 ns rms residual, consistent with the surveyed baseline at the sub-meter level. These results demonstrate real-time, phase-coherent operation of an RDMA-fed Python/CuPy FX backend in the validated four-antenna QUEST configuration.