Syn2Logic: End-to-End Neuromorphic Design Automation

Neuromorphic computing offers a promising path toward energy-efficient, brain-inspired systems, yet the development process remains fragmented. Computational neuroscientists typically work in software-based simulators, while hardware designers manually implement architectures using low-level hardware description languages (HDL). This paper introduces electronic Neuromorphic Design Automation (eNDA) and a prototype framework called Syn2Logic to bridge this gap, allowing researchers to generate custom, synthesizable hardware directly from high-level neural dynamics.

Syn2Logic provides a domain-specific language (DSL) that combines the hierarchical, concurrent nature of hardware description languages with the mathematical syntax used in neuroscience modeling. The framework automates the entire design flow: from lexical analysis and parsing of neural models to high-level hardware optimizations, design-space exploration (such as fixed-point precision and bandwidth tuning), and final synthesis into FPGA bitstreams or ASIC layouts. By decoupling the model definition from the hardware implementation, Syn2Logic allows designers to focus on system-level architecture rather than manual RTL coding.

The authors demonstrate the framework's power through three major applications:

• C. elegans Simulation: They generated the fastest known accelerator for the C. elegans connectome, achieving simulation speeds up to 143,971x faster than state-of-the-art software simulators.
• Sudoku Solver: They created a neuromorphic constraint satisfaction solver that outperforms traditional solvers like CP-SAT and SCIP by nearly two orders of magnitude in speed.
• MNIST Classification: They developed an ultra-efficient MNIST classifier on a 2014-era FPGA, reaching 5.6 million frames-per-second per Watt, outperforming existing neuromorphic architectures in both speed and energy efficiency.