Graphics Processing Units (GPUs) have revolutionized the computing landscape in the past decade and are seen as one of enabling factors in recent breakthroughs in Artificial Intelligence. However, it is very difficult to unlock to full computational power of the GPU. This is because there are many degrees of freedom in GPU programming, and typically only a handful of specific combinations of optimizations and parameter choices result in near-optimal performance. To obtain such highly-efficient kernels it is required to search vast and discontinuous program design spaces, which is infeasible for programmers to do by hand. Moreover, this search process would have to be repeated for different hardware and for different input problems, leading to productivity and sustainability issues with GPU applications. This talk gives a brief introduction to Kernel Tuner, a tool that allows programmers to create tunable applications that can be automatically optimized for any combination of hardware and input problems. Using run-time kernel selection and compilation based on auto-tuning results, programmers can create sustainable applications that can achieve near optimal performance on a wide variety of hardware and inputs.