I will present some work-in-progress/experiments on a high-performance graph-based SDR runtime that differs from popular runtimes such as GNU Radio and FutureSDR in that instead of using circular buffers to connect adjacent blocks, the data travels in packets through closed circuits encompassing multiple blocks of the flowgraph. I show benchmarks comparing GNU Radio 3.10, GNU Radio 4.0, FutureSDR, this new runtime, and a hand-written ad-hoc implementation, for a class of certain simple flowgraphs. These show that popular runtimes often have a high overhead and that there is much room for improvements in performance. This project focuses on ARM Cortex-A53 CPUs, found in AMD RFSoC/MPSoC and other embedded platforms, but most ideas are applicable to other systems, including fast x86_64 machines.