In The Age of Em, I was somewhat confused by the talk of reversible computing, since I assumed that the Laudauer limit was some distant sci-fi thing, probably derived by doing all your computation on the event horizon of a black hole. That we’re only three orders of magnitude away from it was surprising and definitely gives me something to give more consideration to. The future is reversible!
I did a back-of-the-envelope calculation about what a Landauer limit computer would look like to rejiggle my intuitions with respect to this, because “amazing sci-fi future” to “15 years at current rates of progress” is quite an update.
Then, the lower limit is n×kT with n∼10 or n∼100 [...] A current estimate for the number of transistor switches per FLOP is 106.
The peak of human computational ingenuity is of course the games console. When doing something very intensive, the PS5 consumes 200 watts and does 10 teraFLOPs (1013 FLOPs). At the Landauer limit, that power would do 1023 bit erasures per second. The difference is 1010 − 6 orders of magnitude from FLOPs to bit erasure conversion, 1 order of magnitude from inefficiency, 3 orders of magnitude from physical limits, perhaps.
In The Age of Em, I was somewhat confused by the talk of reversible computing, since I assumed that the Laudauer limit was some distant sci-fi thing, probably derived by doing all your computation on the event horizon of a black hole. That we’re only three orders of magnitude away from it was surprising and definitely gives me something to give more consideration to. The future is reversible!
I did a back-of-the-envelope calculation about what a Landauer limit computer would look like to rejiggle my intuitions with respect to this, because “amazing sci-fi future” to “15 years at current rates of progress” is quite an update.
The peak of human computational ingenuity is of course the games console. When doing something very intensive, the PS5 consumes 200 watts and does 10 teraFLOPs (1013 FLOPs). At the Landauer limit, that power would do 1023 bit erasures per second. The difference is 1010 − 6 orders of magnitude from FLOPs to bit erasure conversion, 1 order of magnitude from inefficiency, 3 orders of magnitude from physical limits, perhaps.
Does it take a million bit erasures to conduct a single floating point operation? That seems a bit excessive to me.