This is not true. GPUs are SIMT, but any given thread in those 32 in a warp can also issue SIMD instructions. see vector loads