I think I know what's up
The computer is simultaneously fast and slow
Quantum computing device manufacturer D-Wave is disputing a recently published study that claims the Canadian firm's systems aren't reliably faster than more-conventional computing systems. On Thursday an international team of computer scientists published a widely previewed paper in Science detailing a series of benchmark …
The New Uncertainty Principle, does it work or not, damn it?
D-Wave should be permitted to define their own test, the 'Catch Me If You Can' problem. But they need to be at least 10x faster than anyone else, otherwise FAIL. That would more or less prove that it's working.
Unless they're found cheating with a custom FPGA to solve that one problem. Too easy to be caught, so not likely.
D-Wave should be permitted to define their own test, the 'Catch Me If You Can' problem. But they need to be at least 10x faster than anyone else, otherwise FAIL. That would more or less prove that it's working.
actually, I thought they did; when they announced the two I recall it was something like 3000x faster than the digital computer in their testing. and this is hardly the first time their claims have been disputed: some swiss researchers found similar results a year ago and they were testing with simulated annealing algorithms (which the d-wave is supposedly optimized for).
Google also issued its own response to the Troyer paper, noting that in its tests the firm is getting computational speeds of over 35,500 times that of conventional computing systems.
Yeah, but is that speedup outside of the error ranges? And what problems are they solving? Movie recommendations?
>"And what problems are they solving? Movie recommendations?"
>The usual things. How to download porn faster.
Actually Google does have a class of hard problems that defy algorithmic solutions:
If customer A07f5DE7C4 watches latex granny porn, and customer B79F6E3C8 prefers vinyl doggy style, should Google display ads for crunchy or smooth peanut butter?
And what problems are they solving? Movie recommendations?
In all seriousness, the majority of the problems Google has to solve - from maintaining their huge network of computing resources and workloads, to modeling user preferences - can be transformed into various graph problems. And many of those can be converted into questions of energy distribution in a closed system. As such, annealing approaches, which is what the D-Wave seems to be doing1, should be applicable in many cases, and could offer significant speedups over classical discrete calculation.
Of course that's true of other analog mechanisms, "quantum" or not, as any number of people have noted.
1I take no position on whether it's "true" adiabatic quantum computation (AQC), or "true" quantum annealing, etc, or just a hybrid classical analog-digital machine. I haven't looked at it and I'm not qualified to make that determination anyway.
So basically, bad code causes slowdowns?
No, because there is no code.
It's a "quantum annealing" machine, which is something like a cooling metal, only with configurable strengths between the configuration elements (think analog computer doing neural networks or spin glass annealing), and with more quantum sauce.
I also notice that there are two Google Guys/Gals in the list of authors:
> Troels F. Rønnow - Theoretische Physik, ETH Zurich
> Zhihui Wang - Department of Chemistry, University of Southern California
> Joshua Job - Center for Quantum Information Science and Technology & Department of Physics, University of Southern California
> Sergio Boixo - Google, 150 Main Street, Venice Beach, California
> Sergei V. Isakov - Google, Brandschenkestrasse 110, Zurich
> David Wecker - Quantum Architectures and Computation Group, Microsoft Research, Redmond
> John M. Martinis - Department of Physics, University of California Santa Barbara
> Daniel A. Lidar - Department of Chemistry & Center for Quantum Information Science and Technology & Department of Physics, University of Southern California
> Matthias Troyer - Theoretische Physik, ETH Zurich
I'd rephrase to a machine optimized to handle a certain class of problems, not any one particular problem. What you do is construct the experimental setup so that when the system "relaxes" you end up with the optimal solution to that particular problem. It's all about getting the preconditions right. There's a certain quantum-mechanical sauce (retrieving the 'solution' without collapsing the wave-function) but that's the idea.
Trying to compare it to a system setup with a highly specialized, highly optimized GPGPU system really doesn't do it, in my not so humble opinion. My problem here, and as I recall this was pointed out before, is that you have a hell of a time trying to identify when those problem domains actually overlap between the two methods and for which you should detect a measurable (measure, quantum, uhhh...) set of results that positively (p<.01 I believe) demonstrates that the D-Wave is a quantum computer.
Instead what we end up with is some problem types that demonstrate the D-Wave is ideal for that one particular problem and I mean one, not the whole class of problem, a lot of experiments that show equivalence between classical and quantum, and some for which the D-Wave doesn't come up with a solution. Personally, I am not surprised. First, if this were easy, we'd have done it decades ago. Second, thinking how to properly 'program' the D-Wave would give most anyone migraine's, ala Dr. Feynmann and understanding QM. Lastly, if you don't believe in a personally malignant universe, you will now.
A nice post! I am still waiting for my slides from D-Wave for the curious presentation given at SC13... I was given the nudge not to discuss "certain uses". Go figure, it is only a hobby for me...
However, buying a supercomputer to program your quantum computer (that was the presentation's thrust, though it was mentioned it was part of an iterative process), is marketing genius at a computing conference - no really! Buy 2!!
I have a number of colleagues who really hope this works, but so far the resounding lack of evidence suggests there is much more research to do understanding exactly what the machine they have built *is*. Apparently, if you agree not to use the QC for anything "useful" you can apply online to have a go.....see what I did there?
In another thread I wrote "it's still a really good fridge". A friend of mine who makes deuterium bullets, thought is was a cryogenics research project on first hearing "microKelvin" mentioned....
P.
If you're looking for interesting coincidences (of which I am not certain exist), go read my comment over in Pond Scum:
http://forums.theregister.co.uk/forum/1/2014/06/17/mutation_governs_algaes_quantum_mechanics/
Same class of problem, interesting how nature achieves it, and without any reference to micro-Kelvins.
So, seems to me everyone who buys one sees their benefits but a group of scientists who didn't have to justify the business case for buying it don't think its fast? I know who I believe, and frankly that is the people who laid their jobs on the line recommending the purchase of the machines and use them in real world situations, not a fake lab test that no one will ever run on the system ever.
The D-Wave machine may be perfectly genuine, but surely we all know people who've bought some immensely expensive piece of junk and have then managed to spend long enough promoting it internally before the shiny wore off, to polish the CV, and find a job somewhere else? It may not be that common, but it does happen.
I'd mention two or three cases from my own experience but sadly the details might create enough identification to be libellous; though I can say that two of them involved the purchase of electron microscopes to solve problems that, in fact, could not be tackled in this way.
when the PCI-E bus came out, it seemed too complex and slow.... and everyone said all the other new buses were better, PCI-E was more expensive, sometimes even slower....
then CPU speeds increased, memory got a lot better, and they improved the PCI-E architecture...
I think you know who the winner is today..... :p
when they manage to get a 'pc-standard' basic test going on the D-Wave, then we will get something believable...
Every time I discuss benchmarks with people its sorta like discussing politics. You have to take into account whose doing what. What initial bias they have if its intentional or not. Then you have to take into account who ran the benchmarks and how versed they are in what is being tested.
Most of the time the person running the benchmarks has more experience running one over the others. Specifically if they had an extremely limited amount of time with what they are benchmarking. Then there are in house benchmarks. These will always be skewed somewhat in favor of the companies own software. In part cause they know how to optimize and tweak their software better. Then the other item is left pretty much out of the box or sometimes even unconfigured.
To put it mildly benchmarking and benchmarks mean nothing unless you have decided on the ruleset before hand that is going to be the standard, and constantly check to make sure that the standard has been adhered too. Even then there is a enough of a margin of error that can throw this stuff of.
Ok so deciding once and for all exactly how "quantum" the machine is looks like a hard job, and I even get that we're not quite sure exactly what sort of thing it might be good at. But from a purely pragmatical point of view evaluating its usefulness is not hard at all - take a Hard Problem you're interested in, get the best conventional hardware you can, put it side by side with the D-wave kit and see which gives you more answers per second, by how much, and how accurate / optimal each is. That's it! If you really observe 35000 answers from one while a single one from the other, it really is faster, quantum or not - unless it's 35000 time more expensive too, it's a win. If not, we're just playing "The emperor's new clothes" all over again...
Apart from all the other factors, it's necessary to take into account problem preparation time - an issue identified as far back as the days of one A M Turing, and the driving force behind early programming languages.
It's like the interpreted versus compiled languages issue - if I can develop code to solve a particular problem much faster in, say, Python, than in C++, the slower execution time only matters if I need to solve the same problem many times with different parameters. Your argument would be correct if the class of problem we are talking about is like that. But if the problem is a few-off only, getting 35000 answers from one and one from another may not matter at all.
IE Something that computes (something).
The RL10 rocket engine contains a similar "computer" for mixture ratio control.
Most people today would call it a hardwired analog computer. Operational amplifiers, temperature sensors, no actual ADC/DAC in sight etc.
So what kind of problems do Google get a 1:35000 speed up on?
The overwhelming sense I have, reading about this situation, is that quantum computing is still in an infantile stage. But here we have it being judged in public as if it's a viable product ready for public consumption, when it obviously is not. The enormous Clunk Factor going on at D-Wave is inevitable at this point. To have them shoved into defensive mode over an obviously Not Ready For Prime Time machine is ridiculous. I hope this wrong attitude toward the technology doesn't kill the baby before it has a chance to grow up.
And then there's the possibility that D-Wave really does have a dog that won't hunt. But this clearly is not the time for that judgement.
Considering the QPU is cooled down to millikelvins, in a special room similar to the shielded one used for MRI scanners and shielded from everything including ambient radiation I would say that D-Wave do indeed have a quantum computer.
calling is a "physics experiment" is like calling Fat Man a "Sunshine simulation device"... Doesen't do it justice and D-Wave deserve serious funding whether their system is fully tested or not.
We can't afford to be left behind on this, China has its own version in the pipeline and South Korea even has a prototype under construction.
503 qubits is incredible progress in such a short time even if they are flux ones, ten years ago this would have been science fiction.
Given that ... I would say that D-Wave do indeed have a quantum computer
Although your biggening up is noted, your criteria for QC-ness are frankly lacking in salience.
You are of course free to send your own money to D-Wave (far be it from me to have a negative opinion on that), but please don't go all "we this" and "we that" here.
"And then there's the possibility that D-Wave really does have a dog that won't hunt. But this clearly is not the time for that judgement."
I recall a company that put out some of the first Josephson Junction superconductor components. Made in Nb and LHe cooled.
They were commercial.
Were they "Ready for prime time?"
Probably not.
But they cost their early adoptors a shedload of cash to find that out.