KANs Show Poor Performance on Embedded RISC-V Hardware

Recent research investigated the deployment performance of Kolmogorov-Arnold Networks (KANs) compared to traditional Multi-Layer Perceptrons (MLPs) on embedded RISC-V hardware. The study focused on their use within hard-constrained recurrent physics-informed networks (HRPINNs), where a neural branch learns residual dynamics. The findings indicate that while KANs offer parameter efficiency, this advantage does not translate to deployment costs on scalar embedded cores. KANs were found to be 4.7 to 14.5 times slower and consumed 4.7 to 18.7 times more energy per integration step than MLPs. Furthermore, KANs showed significantly reduced dependability under INT8 quantization, with trajectories diverging much earlier. These results suggest that MLPs remain a more reliable and efficient default choice for embedded HRPINN deployments, unless specific co-design for quantization is implemented for KANs. The study highlights a crucial gap between theoretical parameter efficiency and practical deployment performance for KANs in resource-constrained environments.