It's official, it seems: the D-Wave isn't a “real” quantum computer, but it does handle some classes of problems a lot faster than a classical desktop computer.
That's the result of the first attempt to benchmark the company's adibiatic quantum computer, but it comes with caveats.
But first, some background. D-Wave is a Canadian outfit that says it can sell you a quantum computer right now. The company says its D-Wave Two offers a "512-qubit processor chip ... housed inside a cryogenics system within a 10 square meter shielded room." The company doesn't claim to have a fully quantum computer, but instead to have designed and manufactured "processors required to use quantum effects to compute".
That's understandably contentious, which probably sparked sufficient interest to get this research done. The study was conducted by Catherine McGeoch of Amhurst College in the USA, and Cong Wang of Canada's Simon Fraser University. McGeoch is a 25-year veteran at setting up tests of computing performance and the author of A Guide to Experimental Algorithms.
In a paper to be presented to the ACM conference in Italy on May 15 and published in the New York Times, McGeoch tests the adibiatic quantum computation technique called quantum annealing against various solvers running on desktop computers.
In D-Wave's corner, problems were presented to a 439-qubit instance for solution in one of two modes: quantum annealing for “native instances” that can be solved directly on the quantum hardware; and “Blackbox” mode, a hybrid approach which “alternates heuristic search with hardware queries” (as D-Wave explains, Blackbox is a hardware-software compiler designed to present an abstraction layer so that programmers don't have to try and program their problems directly to the hardware).
McGeoch compared these two approaches to three software solvers: IBM's CPLEX, the open-source METSlib tabu search solver, and the Akmaxsat solver. The tests were run on various problems that fall into the NP-Hard category: quadratic unconstrained binary optimisation, QUBO (a pattern matching problem), the weighted maximum 2-satisfiability problem (W2SAT), and the quadratic assignment problem (QAP).
The performance summary is:
- For QUBO, the quantum annealing solver smacked down the best software solver, running 3,600 times faster than CPLEX for a 439-qubit problem.
- There was no difference between Blackbox and the software solvers for the W2SAT problem.
- For the QAP problem, while the research doesn't mention timeframe, Blackbox outperformed the METSlib tabu solver by finding best solutions in more cases – it turned up 28 solutions for 33 cases, while tabu only found nine.
McGeoch emphasised that it's not yet a “fair” test, however, suggesting that more significant results would be obtained comparing the D-Wave approach to classical hardware that's been optimised for the same set of problems. As is, she notes, a general purpose computer will always be slower than “a device dedicated to solving a specific problem”.
However, the test result – particularly the QUBO test – are being taken as evidence that something quantum-like is happening in the D-Wave, since it converged on the solution almost instantly compared to the classical computer. So it seems that with the right problem – in particular, a problem that maps well onto the quantum hardware – entanglement and superposition are happening.
For D-Wave, the challenge will be to “generalise” its hardware so as not to be confined only to esoteric problems. ®