ReSCom SNN Accelerator Hits 0.05 mJ Per Image on Artix-7

0.05 millijoules per image for MNIST inference on a $100 base FPGA. That’s what ReSCom delivers on a Xilinx Artix-7, and it does it without sacrificing the stability that usually haunts approximate SNN accelerators.

Stochastic arithmetic slashes energy without breaking recurrent dynamics

Spiking Neural Networks are supposed to be the poster child for energy-efficient inference—event-driven, biologically plausible. But hardware realizations have a nasty habit: uncontrolled approximate arithmetic destabilizes recurrent state updates. ReSCom sidesteps that by using stochastic computing only for multiplications in neuron dynamics while keeping addition and subtraction in exact fixed-point. That split lets the accelerator cut hardware complexity where it hurts most (multipliers) without letting errors accumulate across time steps.

Three neuron models, one reconfigurable datapath, runtime knobs

The architecture combines Integrate-and-Fire, Leaky Integrate-and-Fire, and Synaptic neuron models into a unified reconfigurable design. No separate hardware for each—just one datapath that switches behavior at runtime. The key control lever is the stochastic bit-stream length: shorter streams give faster inference and lower energy at the cost of accuracy; longer streams recover accuracy. That’s a concrete, engineer-friendly handle for meeting application constraints on the fly.

0.05 mJ per image at 100 MHz, and room to push further

On MNIST, ReSCom hits 92.80% classification accuracy while burning only 0.05 mJ per image. The abstract claims this beats recent state-of-the-art implementations in energy efficiency, though it doesn't name which ones—I’d like to see a head-to-head on CIFAR or a neuromorphic dataset. Still, the number is striking: that’s roughly the energy you’d burn typing one character on a phone keyboard, and it’s doing full-scan inference on a 28x28 image.

What comes next is obvious: take the same stochastic/exact split to larger datasets and deeper architectures, and see if the energy advantage holds when you need more precision. If it does, ReSCom gives FPGA-based edge inference a concrete path to sub-millijoule SNN inference without the stability headaches.

Source: ReSCom: A Reconfigurable Spiking Neural Network Accelerator Using Stochastic Computing
Domain: arxiv.org