The follow-up nobody noticed
Twelve years after Landauer, another IBM researcher — Charles Bennett — took the principle in an unexpected direction. If erasing information costs energy, Bennett asked, what if we just... don't erase?
He proved, in 1973, that any computation — any algorithm, any function, any logic gate combination — can be implemented without ever erasing a bit. Addition. Multiplication. Matrix math. FFT. Neural network inference. All of it.
The catch
There's always a catch. Bennett's result came with one: reversible computation requires you to keep the intermediate state around. Every temporary result. Every scratch value. Every byproduct of the calculation.
You trade energy for space. The more you're willing to remember, the less you have to erase, the less energy you burn.
The spectrum
Between "erase everything, minimum memory, Landauer cost" and "erase nothing, infinite memory, zero cost" is a full spectrum. You can run a computation in any mode along that spectrum.
- Fully irreversible — standard computing. Minimum memory. Pays Landauer's floor per erasure.
- Partially reversible — some intermediates kept, some erased. Trades space for energy.
- Fully reversible — nothing erased. Theoretical zero energy. Infinite memory.
The engineering question isn't "reversible or irreversible?" It's "which part of this computation should be reversible, and which should pay the erasure cost?"
The Bennett decomposition
Bennett's deeper insight: any operation factors into two parts. An L₀ phase (the reversible part, where information is only rearranged) and an L₂ phase (the erasure part, where information is destroyed).
Take matrix multiply. The broadcast multiply across all input pairs is bijective — every unique pair produces a unique product. That's the L₀ phase. Free. Then the reduction that sums the products into output entries — that's the L₂ phase. That's where you pay Landauer's cost.
Every L₂ operation has an L₀ twin hiding inside it. The stack shows you where.
Why it was ignored
Bennett's result is one of the most important theoretical discoveries in computer science, and almost nobody in the industry acts on it. Why?
Because in 1973, memory was more expensive than energy. Keeping all your intermediates around cost more than paying the erasure. So the field went the other direction — erase aggressively, minimize state, squeeze every operation into the smallest possible footprint.
That tradeoff has completely inverted. Memory is abundant. Energy is expensive. It is now lower-cost to keep state than to erase it. Bennett's result, 53 years later, is finally relevant.
Every computation factors into a reversible phase (L₀, free) and an erasure phase (L₂, pays Landauer's cost). If you can identify which operations in your design are L₀ and which are L₂, you know exactly where to optimize. That's the Bennett decomposition, and the Periodic Stack computes it for every primitive.