Theory 22/Mathematics & Computation /Computing theory/3 min
The Energy Floor Is Data Movement
Landauer's limit is five orders of magnitude below any real chip. The gap is not arithmetic.
The energy cost of computation is set by the bytes moved per operation across the memory hierarchy rather than by arithmetic, so the practical floor is a joules-per-bit curve and efficiency gains come from moving fewer bits, not from doing fewer operations.
01The evidence
- Analyses of data gravity and the energy limits of computation argue that the dominant cost in modern systems is moving data rather than computing on it, and that this changes what an efficiency limit means.Data gravity and energy limits
- Processing-in-memory architectures are being built precisely because moving data between memory and compute dominates energy, and placing compute near memory removes the transfer.Processing in memory
02The tension
Energy efficiency is reported as operations per joule, which encourages algorithmic work that reduces operation counts. But the measured energy of real workloads tracks data traffic, not operation counts, so optimisations that halve operations and double bytes can make things worse. The metric and the physics disagree.
03The theory
Every operation requires operands, and operands live at some level of a hierarchy in which each step away from the arithmetic unit costs orders of magnitude more energy per bit. The total is dominated by the deepest level touched, times the number of bits moved.
That makes the energy of an algorithm a function of its data movement graph: how many bits cross how many boundaries. Arithmetic, at the bottom, is nearly free by comparison. The floor is therefore not a thermodynamic limit on computation but a structural limit on movement.
The theory predicts a universal ladder of joules per bit across technologies, from registers to DRAM to interconnect, and predicts that any algorithm's energy can be estimated from its movement counts before it is run.
It also predicts where the next gains are. Algorithms that restructure data layout to reduce crossings will outperform those that reduce arithmetic, and the advantage should scale with the depth of the hierarchy.
04What this is not
Operation-count metrics assume arithmetic dominates. The Landauer limit is a thermodynamic bound far below practice. This identifies the actual binding constraint as movement, and gives a measurable floor that is many orders above Landauer.
05Predictions
- P1Measured energy should correlate with bytes moved better than with operations performed, across workloads.
- P2Processing-in-memory should deliver a factor predicted from the removed crossings, not from reduced arithmetic.
- P3A joules-per-bit ladder should be measurable and roughly universal across technologies.
06What would refute it
Workloads whose energy tracks arithmetic after movement is controlled would restore the operation-count view.
07What it would change
It would change how efficiency is benchmarked, and make data layout a first-class optimisation target alongside algorithms.
08Evidence base
- Data gravity and the energy limits of computation — arXivhttps://ar5iv.labs.arxiv.org/html/2603.26053
- Enabling practical processing in and near memory — arXivhttps://ar5iv.labs.arxiv.org/html/1905.04376