Weighing one correct answer: 0.35 J on a MacBook, 14.8 J on an RTX 4090
Energy claims in computing are usually stated per hour, per chip or per training run. None of those units answers the question a buyer or a researcher actually has: what does one correct answer cost? I built an open benchmark, energy-per-result, to measure exactly that, and this post reports its first two backends.
The task
The workload is the ground state energy of the hydrogen molecule through VQE: parity mapping, two qubits, run on the Aer simulator. A result counts as correct only if it lands within chemical accuracy, 1.6 milliHartree, of the reference value. Everything else is a wasted joule. The unit reported is joules per correct result, so wrong runs make the number worse, exactly as they should.
The two machines
Backend one is my Apple M1 MacBook Pro, metered with macOS
powermetrics. Backend two is a rented RunPod pod with an
NVIDIA RTX 4090, metered at the GPU board through NVML, running Aer on
GPU. One version note for reproducibility: qiskit-aer-gpu
is at 0.15.x, so Qiskit is pinned below 2 on the GPU pod.
Results
| Backend | Meter | Correct runs | Median energy per correct result |
|---|---|---|---|
| Apple M1 CPU | powermetrics | 10 of 10 | 0.35 J |
| NVIDIA RTX 4090 | NVML, GPU board | 10 of 10 | 14.8 J |
The laptop beats the datacentre GPU by roughly forty times on this workload. That is not a verdict on the 4090, it is a verdict on fit: a two qubit statevector simulation is tiny, and a 450 watt class board spends most of its power waiting for work that never fills it. The interesting question, which the benchmark now makes answerable, is where the crossover sits as the circuit grows.
Why this matters beyond two qubits
Vendors across quantum and classical computing publish energy advantages measured on their own terms: different workloads, different system boundaries, cooling in or out, control electronics in or out. Those claims cannot be compared. A shared unit, joules per correct result on a fixed task with the meter boundary stated, makes them comparable. That is the whole point of the tool.
What comes next
The evidence for every run, including meter screenshots, lives in the
repo under docs/evidence. Next backends in order: a real
QPU through a cloud provider, and an AMD Kria class edge board as a
permanent lab fixture. Each new backend lands as a tagged release, one
per month.
Code and raw data: github.com/tech4biz-yasha/energy-per-result. Questions or corrections: yasha.khandelwal@tech4biz.io.