IBM’s new five-qubit universal quantum computer

[Spinnor]

How significant is this development of a 5 qubit computer by IBM?

See,

"IBM invites users to test its quantum computer"

at http://phys.org/news/2016-05-ibm-users-quantum.html

"How IBM’s new five-qubit universal quantum computer works"

at http://arstechnica.com/science/2016/05/how-ibms-new-five-qubit-universal-quantum-computer-works/

What would be a good math problem for a 5 qubit computer that would really show its power over a regular computer granted it is only 5 qubits computer?

Thanks!

 

[Strilanc]

See,What would be a good math problem for a 5 qubit computer that would really show its power over a regular computer granted it is only 5 qubits computer?

Normal computers can simulate a dozen qubits no problem. (For example, try my in-browser quantum circuit simulator "Quirk".) IBM making 5 qubits available as a service via the internet is purely for the "that's awesome!" response. It's not useful.

That being said, they will benefit from testing the service before they manage to scale up to a useful number of qubits that people would legitimately want to pay to use.

(If the qubits could be kept very far apart, entangled, and coherent, then only two would be sufficient to do interesting coordination tasks. For example, passing spacelike-separated Bell tests. But IBM doesn't have a "fly half of the chip to the other side of the world while magically keeping it coherent" command implemented yet.)

 

[secur]

Spinnor, to use the IBM QC one must sign up ... I think they want serious researchers only?

Strilanc, your simulator blows up unfortunately on both my Windows 10 and android machines. Is it supposed to work on these OS's?

Of course a regular computer can simulate operation of a QC (as long as it doesn't involve long-distance entanglement!) But the one thing it can't simulate perfectly is failure modes. In other words if you want to know exactly what wrong answer IBM's QC will give, you can't use a simulator!

Strilanc it seems to me limiting simulation to only a dozen qubits is unambitious? With modern computers, especially using GPU, one could do more in 50 ms (the maximum time IBM's QC can take to do its program). If a gate operates on all 12 qubits at once, and if all coefficients are complex numbers, the matrix calculation should only take maybe 10 ms (roughly; averaging over multiple cores). But that's very much worst case: most gates are single-qubit or 2-qubit which take essentially no time at all (nanoseconds). 4 or 6 qubit gates might take up to roughly 2.5 ms. Most circuits use not many gates, a dozen or two? Also I'm ignoring many computational tricks, such as assuming coefficients are much simpler than a pair of real numbers; hard-coding standard matrix operations, etc. So, it seems to me one could do quite a bit better than just 12 qubits?

 

[Strilanc]

Strilanc, your simulator blows up unfortunately on both my Windows 10 and android machines. Is it supposed to work on these OS's?

Does it show the "An error happened :(" prompt? If so, you can use the links to report the issue or just copy paste the information here. webgl is quite finicky, so there's a lot of "it works by accident on my machine" happening.

Strilanc it seems to me limiting simulation to only a dozen qubits is unambitious? With modern computers, especially using GPU, one could do more in 50 ms (the maximum time IBM's QC can take to do its program). If a gate operates on all 12 qubits at once, and if all coefficients are complex numbers, the matrix calculation should only take maybe 10 ms (roughly; averaging over multiple cores). But that's very much worst case: most gates are single-qubit or 2-qubit which take essentially no time at all (nanoseconds). 4 or 6 qubit gates might take up to roughly 2.5 ms. Most circuits use not many gates, a dozen or two? Also I'm ignoring many computational tricks, such as assuming coefficients are much simpler than a pair of real numbers; hard-coding standard matrix operations, etc. So, it seems to me one could do quite a bit better than just 12 qubits?

Actually, it's much better than that. On my work machine Quirk stays responsive with 16 qubits and hundreds of gates being applied sequentially (i.e. it recomputes the entire output superposition from scratch and breaks down the marginal state of each qubit). Quirk updates in real time as you drag gates around. It's not smooth with that many qubits and operations, but it's usable. GPUs are amazing.

I would allow more qubits, but it's a design goal for Quirk to feel like you're interacting with the circuit instead of giving a "compute!" command and waiting for the result. So 100ms is a hard limit on how long it can take to compute and draw a frame. (Google's quantum computing playground uses the same compute-with-webgl approach, but doesn't need to be quite so interactive, so they go up to 22 qubits. And that's probably more of a memory limitation than a time limitation.)

I didn't intend for the "dozen qubits" quip to be taken as an upper limit.

 

[secur]

Wouldn't be surprised if GPU is the problem. I'm extremely familiar with SSE and AVX but still haven't gotten into GPU, although I've intended to for years. Main problem: they're so non-standard that supporting different machines can be major headache (as, perhaps, you're learning now with Quirk :-) If user doesn't have AVX2 I can easily drop back to SSE 4.1 (everyone has that); not so simple with custom vector routines on GPU.

I would like to get Quirk running but don't want to hijack this thread. I can email from Android, but Windows 10 still not set up (it's better than my previous 8.1 but many upgrade problems). So here's the errors from that machine. After this perhaps I'll start a new thread if want to discuss it further, or go to GitHub.

BROWSER USER AGENT
Mozilla/5.0 (Windows NT 10.0; WOW64; Trident/7.0; .NET4.0C; .NET4.0E; .NET CLR 2.0.50727; .NET CLR 3.0.30729; .NET CLR 3.5.30729; LCJB; rv:11.0) like Gecko

ERROR OBJECT
TypeError: Object doesn't support property or method '__$333167720$481339502$__'

ERROR LOCATION
TypeError: Object doesn't support property or method '__$333167720$481339502$__'
at qubitDensityMatrixToBlochVector (src.min.js:1050:20)
at paintBlochSphere (src.min.js:1259:118)
at drawOutputDisplays (src.min.js:1518:93)
at paint (src.min.js:1487:1)
at paint (src.min.js:1544:108)
at F (src.min.js:948:33)
at trigger (src.min.js:386:63)
at E (src.min.js:944:123)
at G (src.min.js:949:62)
at Anonymous function (src.min.js:955:1)

 

[Strilanc]

Issues created:

https://github.com/Strilanc/Quirk/issues/111

https://github.com/Strilanc/Quirk/issues/112

 

[secur]

What would be a good math problem for a 5 qubit computer that would really show its power over a regular computer granted it is only 5 qubits computer?

Perhaps it's worth mentioning that all those standard "Alice, Bob and Eve" QC examples can be done with less than five qubits: quantum teleportation, dense coding, pseudo-telepathy, cryptography. You just have to pretend that some of the entangled qubits are separated by a few light-years. A good exercise, even though it doesn't "really show its power". I wonder what the IBM QC's rate of success would be.