When I was optimizing GPU kernels a few years back, Nvidia's own kernels were getting those last few percent of performance by making use of hardware-specific f

When I was optimizing GPU kernels a few years back, Nvidia's own kernels were getting those last few percent of performance by making use of hardware-specific features (I remember operand caches being one) that are not available through PTX.