Hey, MD here... You are right that these tradeoffs are difficult to engineer.
However, it's easier to make it work on top of a more radical shift in hardware. At the basic level, we are just using simple random matrix-vector mults. Of course, the photonics or other analog low-energy approaches have to win in the market for this operation, and that will be tested empirically (though there has been a ton of investment into this kind of processing going analog as we discuss in the paper).
However, it's easier to make it work on top of a more radical shift in hardware. At the basic level, we are just using simple random matrix-vector mults. Of course, the photonics or other analog low-energy approaches have to win in the market for this operation, and that will be tested empirically (though there has been a ton of investment into this kind of processing going analog as we discuss in the paper).