Workshop: Twelfth International Workshop on Heterogeneous High-Performance Reconfigurable Computing (H2RC 2026)
Authors: Linus Jungemann, Bjarne Wintermann, and Heinrich Riebler (Paderborn University, Paderborn Center for Parallel Computing); Felix Jentzsch and Christoph Berganski (Paderborn University); and Christian Plessl (Paderborn University, Paderborn Center for Parallel Computing)
Abstract: In dataflow architectures, first-in-first-out (FIFO) buffers decouple pipeline stages
working at different rates. Appropriately sizing FIFOs prevents deadlocks and sustains throughput, yet remains a difficult problem for automated deep neural network (DNN)-to-accelerator compilers. As DNNs increase in size and complexity, existing techniques suffer from excessive runtime, inaccurate results, or unacceptable manual modeling effort.
We present a FIFO sizing approach based on iterative simulation of mixed RTL/HLS black-box
designs. Two innovations manage the cost of cycle-accurate simulation: a resource-aware search algorithm that operates at memory-primitive granularity rather than
raw depth, minimizing simulation rounds; and a scalable, MPI-capable, massively layer-parallel simulation of only the control path, greatly accelerating individual runs.
Integrated into the FINN DNN-to-accelerator compiler and evaluated on a range of DNNs
against three existing techniques, the approach delivers near-optimal FIFO depths with guaranteed performance at an average of 2.88x the speed of the best competing approach offering the
same guarantees.
Back to Twelfth International Workshop on Heterogeneous High-Performance Reconfigurable Computing (H2RC 2026) Archive Listing
Back to Full Workshop Archive Listing