The point stands, the vast majority of workloads are unsuited for GPUs, either because they are full of divergent branches, or because the data transfer and synchronization overheads would cancel any performance gains.