In theory, quantum computers could be much more efficient at some kinds of tasks, which could be potentially disruptive in applications areas such as cryptography. But you know: in theory, theory and practice are the same, but in practice, they are not. So it’s interesting to find applications where quantum computing might possibly be useful and cost-effective in the near(ish) future.
So let’s talk about cell-phone towers. A 5G base station contains a phased-array antenna, that is, a 2-d array of small antennas that can be aimed in a particular direction (for sending or receiving) by adjusting the phase of their signal. And with the superposition principle, this phased-array antenna can simultaneously aim one signal at your iPhone here, while aiming another signal at that Android phone over there, and hundreds more devices all at once. To do that the base station must compute a difficult (NP-complete) combinatorial optimization problem every few milliseconds, as the phones move around. This is called “Massive MIMO,” Massive Multiple-Input-Multiple-Output communication.
Why bother? Well, in dense and expensive cities, it’s so expensive to construct another cell-phone tower that it’s worth the cost of continuous heavy computation to optimize the number of simultaneous connections that each tower can handle. That cost is mostly in the electricity it takes to run a big high-powered compute server for the base station–both in dollars and in carbon footprint.
Recently my colleague Kyle Jamieson and his student Srikar Kasi published a couple of papers demonstrating that a commercially available Quantum Annealing computer, the D-Wave, can solve toy-scale MIMO antenna optimization problems. Unlike “gate model” quantum computers, which implement Von Neumann machines that can compute in a superposition of states, quantum annealing finds optimal or near-optimal solutions to node labeling on a weighted undirected graph. You might think that this is not very general, but plenty of real-world optimization problems can be encoded into this graph problem. And, more to the point: small-to-medium scale quantum annealing computers are practical and commercially available now, whereas gate-model quantum computers are not yet practical except at tiny scale.
But still, quantum annealing computers (such as the D-Wave) are expensive, and require special refrigeration units to get them down to 15 milliKelvin. Could there really be an application where it’s more cost-effective to buy one of those, instead of just a rack full of Arm servers?
Massive MIMO, the antenna-aiming problem, is a good candidate: it demands that the same specialized NP-complete problem be solved every millisecond; and improving the optimization means that more cell-phone conversations can be handled by the same expensive cell tower.
A new paper by Srikar Kasi and colleagues at Princeton, InterDigital, and University College London tries to answer the question, “based on current trends in the improvement of commercially shipped quantum annealing hardware, how many years in the future will be it cheaper to run a quantum annealer than to run a rack full of conventional servers?” And the answer is, maybe in the range 2027-2036 the two solutions will be equally cost-effective for medium-large cellular base stations. And maybe after 2036 the quantum annealer will consume 45% less electricity (be 45% cheaper to run) than conventional computers for large cellular base stations.
That’s pretty exciting. Because up to now I thought that quantum computing is the technology of the future and always will be. But 2036 is only fifteen years away, and maybe we’ll really see this happen.
Additional remark: Quantum simulators are another kind of special-purpose (not gate-model) soon-or-now practical quantum computer. What I find interesting about Massive MIMO is that the quantum computer is solving a combinatorial optimization problem that is not directly about quantum mechanics.
Note the kinds of problems that you are seeing successes with: they are ones involving physical simulation, not number-theoretics.
Remember the old analog computers from the early 20th century? The ones that could solve physical-simulation problems more quickly than the early digital computers? But they also did so with less accuracy. And they were useless for number-theoretic problems, which could only be handled by digital computers.
In other words, “quantum” computers are just a new kind of analog computer.