Nvidia states that a diffusion model can achieve up to 15x speedup in processing. The improvement allows entire blocks of text to be generated simultaneously.