Performance overview of GPU-accelerated butterfly matched filtering on a quad-GPU node. Performance of clFFT in CSP/Out-of-Place as a function of array size N for two batch sizes measured by allocation M in global memory (left panel). Array sizes exceeding 2¹² cause a drop in performance as matrix transpose exceeds 32 kB of local memory, limited now by 1TB/s bandwidth to Global Memory in HBM2. We note that clButterfly uses a default size N = 2¹⁷ (32 s of data at 4096Hz sampling rate) with batch size over n = 128 per frame of 4096 s. A quad-GPU node hereby features a maximum of about 5 teraFLOPs, circumventing limitations of PCIe bandwidth optimized by pre- and post-callback functions (van Putten 2017). The throughput of this quad-GPU node is shown up to six times real-time over a template bank normalized to size 2²⁰ (right panel). GPU-CPU communication passes through a post-callback function, limiting data-transfer to tails exceeding a threshold κ = 2 in normalized butterfly matched filtering output, effectively 0.1% of all matched filtering results. Throughput peaks in multi-frame analysis at a cross-correlation rate of about 200kHz.