Answers are grounded in this paper and cite their sources.
Systems-on-chip that combine a multicore processor with programmable logic (PL), the FPGA part of the chip, allow computation to be offloaded to an accelerator built in the PL. Whether this pays off depends on the workload and on the cost of moving data. On an AMD Kria KV260, we measure the offloading boundary, i.e., the problem sizes and computational intensities beyond which the accelerator outperforms the processor, and test whether a model predicts it. A synthetic microbenchmark, whose computation per element is set by a parameter K, runs on one and four Arm Cortex-A53 cores and on two accelerators generated by high-level synthesis. A three-parameter model reproduced the accelerator times within 0.04% and predicts that, at 100 MHz, the PL must process more than about five elements in parallel to outperform one core and about 21 to outperform four. A four-lane accelerator confirmed the predicted computation rate but, without burst memory transfers, was 3.3–77 times slower than one core. Under controlled measurement conditions, no accelerator outperformed four cores in time or energy: the better one was 13–22 times slower and used about 10–13 times more energy. Memory-access overheads and the processor baseline are thus first-order design parameters.
No takes yet. Share an insight, caveat, or question.
Sabatini et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: