QBF-Based Synthesis of Optimal Word-Splitting in Approximate Multi-Level Storage Cells

In applications such as multimedia that tolerate imprecise results, approximate computing techniques can sacrifice precision to save power or time. One aspect of approximate computing is imprecise storage in multi-level cells (MLCs). Computer words that are too large for a single MLC must be distributed across multiple approximate MLCs. The wordlevel imprecision depends on how the words are split across the MLCs. This work gives an automated synthesis approach for splitting words across MLCs. Given bounds on the imprecision of individual MLCs, the technique synthesizes solutions for splitting words across cells to minimize the worstcase imprecision at the word level. The technique is based on quantified Boolean formula solving, within an overall optimization loop. Worst-case word-level error is shown to vary by over an order of magnitude across otherwise comparable word-splitting alternatives.