EigenOrbit Study Finds No Single Best Quantum Backend
Here’s the thing nobody running a quantum demo wants printed on the slide. There is no best backend. Not IBM’s. Not IQM’s. Not whatever simulator you’ve been treating as your house default since grad school. A new study from the EigenOrbit team ran the same quantum workloads across twelve execution providers and found the winner kept changing depending on what you asked it to do.
The study is titled eigenkit: A Provider-Neutral Quantum Computing Library and is an arXiv preprint. It introduces an open-source Python library called eigenkit and then uses that library to do something the field talks about constantly but rarely tests cleanly: take one algorithm, run it everywhere, and see what happens.
What the study measured, and how big
The point of eigenkit is to pull the algorithm apart from the machine running it. You define a circuit once. Then you fire it at whatever backend you want through one shared interface, no rewriting the workflow for each provider’s SDK. That separation is what makes the comparison fair, because every backend is handed the identical thing.
The scope, straight from the study:
- Twelve execution providers total, all measured (eight simulator backends, four hardware routes)
- Two named quantum processors in the hardware experiments, measured: IBM Heron and IQM Emerald
- Backend-independent gradient computation plus hooks into NumPy, PyTorch and scikit-learn, a built capability
- Simulator performance that shifts with circuit size, structure, batching and workload, a measured behavior across tasks
The number that moves
On simulators, the result is almost mundane until you sit with it. A simulator that’s great at one task can be noticeably worse at another. Circuit size, how the gates are shaped, whether you batch, all of it swings the outcome. So your trusty simulator isn’t a constant, but a mood.
Hardware is where it gets interesting. On IBM Heron and IQM Emerald, the better processor for a given algorithm didn’t come down to raw specs. It came down to how closely the circuit’s topology matched the chip’s physical connectivity. Think of it like parallel parking. A great driver in a tiny car and a mediocre driver in a tank will both struggle, but the one whose vehicle fits the space wins, and the space keeps changing. Match the circuit’s wiring to the hardware’s wiring and you get cleaner execution. Mismatch it and you pay for every qubit you have to shuffle around.
No quantum advantage is claimed anywhere in this. That’s not what the work is. It’s benchmarking and tooling, and it’s honest about being exactly that.
Why cross-backend tests usually lie
Comparisons like this fail quietly all the time. You run algorithm A on provider X, algorithm B on provider Y, and declare X faster. Except you never ran the same thing on both.
Why?
Because rebuilding the workflow for each provider is annoying, and annoyed developers cut corners. Different circuit transpilation, different error-mitigation settings, different shot counts creep in, and suddenly you’re comparing two things that were never the same experiment. eigenkit’s design closes that gap by refusing to let the workload change between runs. The study still doesn’t publish every seed and batch size and SDK version you’d want for a full reproduction, so the validity ends where those details end. But the core comparison holds because the input was pinned.
The researchers are upfront that the goal isn’t to flatten the differences between platforms. It’s to make those differences measurable before you commit a workload to one of them.
What you actually do with this
If you’re picking where to run something, the takeaway is small and useful. Benchmark the workload you have, on the backends you can reach, before you pick. Backend choice is a decision you make per job, not a property your algorithm carries around forever.
One PQC note, since quantum software and cryptography keep colliding. Tooling that lets you test circuits across hardware is separate work from the cost-cutting happening on the security side, where developers recently pushed a quantum-safe Bitcoin transaction below $70. Different problem, same impulse to measure instead of promise, the way regulators pushing the NIST 2030 and 2035 PQC deadlines want the whole field to behave.
eigenkit is on GitHub. Go run your own circuit on all twelve and see which one wins for you. It won’t be the same as mine. For more on quantum software and benchmarking, keep reading QNews.