Synchronous stochastic gradient descent (SGD) is an approach to parallelizing SGD. Although synchronous SGD parallelizes well in an idealized computing environment, it fails to realize its potential acceleration in many practical settings. One cause is slow workers, termed ``stragglers". In many approaches to straggler mitigation, work completed by stragglers, while only partial, is discarded completely. In this poster, we propose an approach to parallelizing synchronous SGD that exploits the work completed by both slow and fast workers to reduce convergence time. We provide a convergence analysis for our approach. Our numerical results confirm an improvement of several factors of magnitude in comparison to existing methods.