sech1: might be a stupid question, but if they are programmable why can't they do floating point calculations? The way it was explained to me was that DPUs (or at least new generation ones) use far less power than alternate CPU/GPU racks, and can also be better parallelised (at least that was my understanding of the explanation). It was explained by a guy from Nvidia to me at a conference.

that will be 10 times slower than the real FP, and they probably don't have enough cache anyway