Reproduce “CircuitsDNA: Discovering Unconventional Multi-Accuracy Arithmetic Circuits via Evolutionary Synthesis” on OpenResearch

About this paper

Ruichen Qi, Junyi Luo, Xinting Jiang, Quan Cheng, Gregory Kielian, Ben Laurie, Dennis Sylvester, Mehdi Saligane

CircuitsDNA is an evolutionary synthesis framework developed by researchers at Brown University, Google, and the University of Michigan that automatically generates configurable arithmetic circuits, enabling dynamic trade-offs between accuracy and efficiency by supporting multiple error modes within a single hardware instance, demonstrating up to a 56% reduction in Area " Times " Power product for 8-bit multipliers in deep neural network workloads.

Key insights

• A single hardware netlist can be evolved to concurrently satisfy distinct, multiple runtime-selectable error constraints for different accuracy modes, allowing dynamic PPA-accuracy trade-offs within one physical circuit.
• Integrating formal verification into evolutionary circuit synthesis can be made efficient by combining time-bounded SAT checks for fast rejections with an asynchronous fallback mechanism for complex, time-consuming proofs, preventing search stalls.
• Dynamically adapting the probabilities of mutation operators based on their historical effectiveness (via credit-driven Exponentially Weighted Average) significantly accelerates the convergence of evolutionary search in complex combinatorial design spaces.
• Unconventional approximate circuit structures can be automatically discovered through logic-level evolutionary synthesis, potentially yielding greater PPA benefits than template-based or manually designed configurable circuits.
• The effectiveness of approximate circuits varies across different AI workloads and layers, underscoring the value of runtime-configurable units that can adapt to specific error tolerance requirements for optimal energy efficiency.

Summary by alphaXiv.