ArKANe: Accelerating Kolmogorov–Arnold Networks on Reconfigurable Spatial Architectures

Kolmogorov-Arnold networks (KAN) are promising for explainable machine learning on embedded devices. Although KANs require fewer model parameters and are smaller than traditional deep neural networks, their applicability is limited by the complexity of the B-spline functions used at every neuron. Hence, this work presents acceleration techniques for KANs on dataflow-processing-based accelerators implemented in AMD-Xilinx Versal ASoCs. Our experiments on the VCK-190 devices demonstrate that for large batch sizes, the average B-spline computation on systolic and wavefront architectures outperforms standard CPU implementation by 11x and 60x. On the other hand, wavefront accelerators demonstrate more than 3x improvement over GPUs (RTX 3090) in terms of energy consumption. Thus, this work opens new opportunities for spatial accelerators in non-conventional machine learning algorithms on embedded systems. This work’s code and experimental artifacts are available at https://github.com/SPIRE-GMU/SPIRE-ARKANE.