Where the free lunch goes

Reversible computing has an unfair reputation as a free lunch. The pitch is irresistible: if no information is erased, no energy need be dissipated, and the Landauer floor simply does not apply. Every few years someone rediscovers this and concludes the energy problem is a matter of insufficient cleverness.

The catch is that reversibility does not remove the cost. It converts it into a different currency, and the exchange rate is brutal.

To approach the reversible limit you have to run the computation slowly enough that the system stays near equilibrium throughout. Dissipation falls roughly in proportion to how long you are willing to take. Halving the energy means roughly doubling the time. That is a real trade and sometimes a good one, but it is a trade, not a saving — and it is exactly the wrong trade for inference, where latency is most of the product.

There is a second cost that gets discussed even less. A reversible computation cannot throw anything away, which means every intermediate result has to be kept until it can be uncomputed. The garbage does not vanish; it accumulates, and then has to be unwound. You have traded energy for time and for memory, and memory is where a large part of the energy was going in the first place.

You can have it fast, you can have it cheap, or you can have it small. Reversibility does not repeal that; it lets you choose differently.

None of which makes it uninteresting. It makes it a design space rather than an escape hatch. The question worth asking is not “can we compute without dissipating” but “which parts of this workload are latency-insensitive enough to be worth slowing down?”

My suspicion is that the answer is: more than we think, but not the part everyone is looking at. Training is throughput-bound and could plausibly absorb a large slowdown in exchange for a large energy win. Inference is latency-bound and cannot. If reversibility lands anywhere it lands upstream, which is a much less glamorous claim than the one usually made for it.