SA-Kura: First Digital Accelerator for Kuramoto Diffusion Sampling

SA-Kura has been unveiled by researchers as the inaugural digital systolic-array accelerator tailored for locally coupled Kuramoto drift in diffusion sampling. While diffusion inference poses high costs for edge deployment, current accelerators primarily target score networks due to the simplicity of standard drift scaling. In contrast, Kuramoto orientation diffusion enhances sampling efficiency through locally coupled phase interactions, yet it introduces a hardware limitation: a center-dependent nonlinear 5x5 stencil that must be evaluated at each reverse step. This kernel struggles with traditional CNN accelerators and matrix-oriented engines. SA-Kura innovatively reformulates pair-wise sinusoidal coupling into neighbor accumulation that is independent of the center phase, followed by a center-dependent multiply-subtract operation, thus eliminating in-PE transcendental units and facilitating regular systolic execution with register-level reuse. This research was published on arXiv.