For years graphics co-processing units (GPUs) have been growing in performance capabilities<br>faster than CPUs, and have recently surpassed them. Simultaneously the hardware and<br>corresponding APIs have grown more generally applicable to a wide range of computational<br>tasks. This has fostered the use of GPUs for general programming.<br>In this abstract we will describe our experiences monitoring, testing and ultimately using<br>GPUs to run various seismic imaging kernels. The performance improvement of GPUs over<br>CPUs varies with algorithm but is high enough so that we've purchased a substantial GPUbased<br>cluster. In our workshop talk, we will discuss our production experiences running code<br>on this system.