There are a few ways to access real quantum computers — often for free — over the Internet. However, most of these are previous-generation machines that have limited capabilities. Great for learning, perhaps, but not something you could do anything practical with. Xanadu, however, has announced what they claim to be a computer capable of reaching quantum advantage that is free for anyone to use, within limits. Borealis — the computer in question — uses photonic states and has the capability of working with over 216 squeezed-state qubits.
The company is selling time on the computer, but the free tier includes 5 million free shots on Borealis and 10 million shots on an earlier series of quantum computers. You can also buy pay-as-you go service for about $100 per million shots on Borealis.
While a few million shots may sound like a lot, we noticed that the quickstart demo consumes 10,000 shots and that’s presumably something simple. That’s still about 500 runs of that on Borealis — not bad for free on a state-of-the-art quantum computer. You will be wanting to debug with a simulator, though.
We presume the developers are Beatles fans given that you use software called Penny Lane and Strawberry Fields to access the machines. Your job is controlled by Python and there is a cloud simulator to save your shots.
We won’t pretend to understand all there is about squeezed light qubits and the Borealis architecture. But you can get some general practice in our series on quantum computing. Or there are a few lectures around including one that aims at different levels of experience.
I am a scientist trying to use “Quantum Computing” (closer to quantum information theory in my case) for my work. My view is usually quite different from a “computer scientist” point of view, since I am more interested in using a quantum computer on quantum data (as opposed to quantum computer on classical data). For my purposes, this is a much more exciting news, since Xanadu’s implementation (photonic quantum computing) allows continuous variable states. Such “Qumodes” in photonic systems have significant advantages in certain types of neural networks, and certain types of Hamiltonian simulation (beyond İsing model). One can implement a quantum variational autoencoder using CV with much less friction than discrete qubit methods, due to embedding, and you have cool “gates” like Doktorov operator, which is quite handy for theoretical chemistry etc.
Strawberry fields offer some very nice tutorials on how to use their machines for neural networks (i.e. https://strawberryfields.ai/photonics/demos/run_quantum_neural_network.html).
In short, I really do hope more computer scientists adopt this technology, and maybe the financial sector gets interested in it as much as they got interested in the super conducting discrete qubit machines, so that it becomes a viable method of doing my work in the medium-long term.
(BTW, if you are interested in researching “white hat” hacks on the quantum internet, and certain quantum key distribution implementations, I recommend looking at “Gaussian cloning” https://strawberryfields.ai/photonics/demos/run_gaussian_cloning.html, i.e. no-cloning theorem states you can not perfectly copy a quantum state, it does not say you can not copy %66 of the quantum state, then the rest is a play of who does statistics better.)
There are other differences between their approach and what is typically first taught in Quantum computation: qumodes (“continuous” ) instead of qubits, but also irreversible instead of reversible, measurement-based instead of gate-based, clusterstate. They start with a highly entangled clusterstate then perform measurements that remove entanglement relationships rather that starting with unentangled qubits or qumodes and building the entanglement.
Metaphorically, the more common approach is that of sculpting by means of adding and molding clay and theirs is that of subtracting with chisel against rock. Both approaches work as a means of performing the computation. However, if I remember correctly, their software permits you to write your code in gate-based fashion and translates into measurement-based, which is actually fairly standard albeit technical.
I also remember that, while their current hardware requires supercooling, I believe for the detectors, they are confident that it will be possible to upgrade using existing technology so that their systems can run under normal room temperature conditions. In time this could mean a quantum computer in your palm.
They are definitely one of the companies that I am more hopeful about.
It is true that the circuit model is not the only way to harness quantum information theory for practical computing tasks, the most promising alternatives being the “one-way quantum computing” more preferred in Neutral atom, or ion trap machines like in Pasqal and my personal favorite cluster state computing. However I would like to emphasize that “their implementation” (i.e. photonic quantum computer) is perfectly capable of circuit model, and universal computing. Please see KLM protocol (the one uses non-linear Kerr gates) LOQC protocol as the most famous examples. These protocols provide you with discrete qubits, completely reversible gate set (i.e. all operators are perfectly Hermitian unitaries), and they provide a set, that allows universal computation.
Xanadu X8 chips use GKP qubits (https://arxiv.org/pdf/quant-ph/0008040.pdf) (https://arxiv.org/pdf/2010.02905.pdf) and they provide a circuit model python interface by transpiling the circuit to the universal unitary in the chip. Hence the same need for ancillary qubits and etc. are still there in order to do computations in that manner. Taking the risk of oversimplification, “Just like the IBM Q, but with a different gate set”. There are a number of disadvantages over IBM Q if you plan to use it in this manner though.
The new chip uses a time domain protocol. I think i should leave this to them to explain: https://strawberryfields.ai/photonics/demos/run_time_domain.html
In any case, I agree with you that it seems like a waste to use their device for discrete qubit scenarios, but if you insist, you can.
Who knows? Maybe in a couple of years someone will solve the problem of room temperature single photon detection, and no one will care it is a waste to use that machine like that, since it is ridiculously cheaper compared to other architectures (did you know that coolers for SC computers are considered a strategic resource, so their resale is strictly controlled?).