Enhancing Speedups for FPGA Accelerated SPICE through Frequency Scaling and Precision Reduction

Frequency scaling and precision reduction optimization of an FPGA accelerated SPICE circuit simulator can enhance performance by 1.5x while lowering implementation cost by 15 -- 20%. This is possible due the inherent fault tolerant capabilities of SPICE that can naturally drive simulator convergence even in presence of arithmetic errors due to frequency scaling and precision reduction. We quantify the impact of these transformations on SPICE by analyzing the resulting convergence residue and runtime. To explain the impact of our optimizations, we develop an empirical error model derived from in-situ frequency scaling experiments and build analytical models of rounding and truncation errors using Gappa-based numerical analysis. Across a range of benchmark SPICE circuits, we are able to tolerate to bit-level fault rates of 10--4 (frequency scaling) and manage up to 8-bit loss in least-significant digits (precision reduction) without compromising SPICE convergence quality while delivering speedups.