[KayEss]

You wouldn't need to render the whole frames. You'd get even better results from randomly sampling over a number of frames to contribute to nearby pixels. In effect you're trading less work for more noise, and noise is kind of what you want here (so long as it accurately reflects what is happening).

So, you could render 16 samples, but only do 1/4 of the pixels on each so you'd get better motion blur for only a small increase in work over doing 4 temporal samples. The extra work corresponds to a bit of bookkeeping and the tweening of 16 frames instead of 4 and you'd still be rendering the same number of samples overall.