Realistic simulation of cardiac electrophysiology requires both high resolution and computationally expensive models of membrane dynamics. Optimization of membrane models can therefore have a large impact on time, hardware, and energy usage. We tested both CPU-based and GPU-based optimization techniques for a human heart model with Ten Tusscher-Panfilov 2006 dynamics. Compared to a multithreaded code running on 64 CPU cores, the tested NVIDIA Tesla P100 GPU proved about 3 times faster. Effective use of the CPU's SIMD capabilities allowed a similar performance gain. GPU performance was bounded by the data transfer rate between GPU and main memory. Optimal SIMD use required explicit vectorization and an adapted data structure. We conclude that on mixed CPU-GPU systems the best results are obtained by optimizing both CPU and GPU code and using a runtime system that balances CPU and GPU load.