Why Amazon Might Become the Largest Quantum Consumer (nextplatform.com) 39
An anonymous reader shares a report:These are still early days for quantum computing, far too soon to talk about domain-specific quantum systems. But if there are areas hungrier than ever for what quantum is best at -- dense optimization problems at scale -- the future cannot arrive fast enough. More specifically, the golden grail for quantum computing -- the "traveling salesman" problem -- could revolutionize the transportation industry in particular, in addition to the world's largest retailers dependent on accurate shipping data. Quantum capabilities in this arena are so critical that the first production quantum systems at scale could be purpose-designed and optimized simply for this type of problem. While these days we don't think of Amazon's delivery aspects much since the carriers are so often the focus, the combined capability of vast search coupled with near-real-time delivery dates matched to location took Amazon years to get right -- and was a billion-plus dollar effort in compute time.
Peter Chapman says "infinite compute" can be brought to bear to refine the entire process that happens the moment you search for "USB drive" on Amazon, confirm your shipping location, and select only products that arrive tomorrow. The density of calculations required -- pulling from warehouse availability to planes, trains, and automobiles and their various routes through your own hometown -- is staggering. "It's the ultimate traveling salesman problem," he laughs. Chapman should know what this takes because he led the development of many of the technologies that became the fast, reliable Amazon Prime service. As director of engineering, his team of 240 engineers took Amazon from requiring customers to search and select a product and wait until checkout to find out how long delivery would take. "That meant a lot of abandoned carts and a bad user experience," he says.
With global products, shipping routes, customers, carriers, product availability and warehouse locations, the order was so tall, it took rearchitecting Amazon infrastructure to do it at reasonable enough scale. "There is a practical limit to the computational resources you can apply to this, even at Amazon. We could easily consume 100x the compute but Amazon couldn't afford it," Chapman says. "There is infinite need for compute for this problem so we had to find the right tradeoffs in optimization and find what you can get for a certain amount of money spent -- and we're talking billions here. Our goal was to make sure it wasn't $20 billion." He adds that the cost of these systems were growing faster than the top line of Amazon's sales.
Peter Chapman says "infinite compute" can be brought to bear to refine the entire process that happens the moment you search for "USB drive" on Amazon, confirm your shipping location, and select only products that arrive tomorrow. The density of calculations required -- pulling from warehouse availability to planes, trains, and automobiles and their various routes through your own hometown -- is staggering. "It's the ultimate traveling salesman problem," he laughs. Chapman should know what this takes because he led the development of many of the technologies that became the fast, reliable Amazon Prime service. As director of engineering, his team of 240 engineers took Amazon from requiring customers to search and select a product and wait until checkout to find out how long delivery would take. "That meant a lot of abandoned carts and a bad user experience," he says.
With global products, shipping routes, customers, carriers, product availability and warehouse locations, the order was so tall, it took rearchitecting Amazon infrastructure to do it at reasonable enough scale. "There is a practical limit to the computational resources you can apply to this, even at Amazon. We could easily consume 100x the compute but Amazon couldn't afford it," Chapman says. "There is infinite need for compute for this problem so we had to find the right tradeoffs in optimization and find what you can get for a certain amount of money spent -- and we're talking billions here. Our goal was to make sure it wasn't $20 billion." He adds that the cost of these systems were growing faster than the top line of Amazon's sales.
Diminishing returns? (Score:4, Interesting)
Re: (Score:2)
Same could be said about autonomous vehicles. An even bigger demand for quantum than logistics.
Re:Diminishing returns? (Score:5, Informative)
Speaking of diminishing returns, although it is true that the absolute shortest route among all possible routes is an NP hard problem, there are good algorithms to find routes that are within an absolute factor (call it "epsilon") of the shortest route for any value of epsilon. That is, you can get arbitrarily close to the TSP shortest route at reasonable computational effort (you get decide on the tradeoff of effort and closeness). It just won't be the shortest of all possible routes.
What you say? Your journey of a 1000 miles could be one step shorter? Why do you care?
Re: (Score:2)
Once you've reduced the measurement to steps, you're already going to be able to get that close with a reasonable computation in most cases. There might only be a 0.1% chance that the shortest possible route is a full step shorter.
To even keep the problem hard, you need to be focused on a continuous measurement like time, and be able to potentially get there nanoseconds earlier.
Re: Diminishing returns? (Score:2)
Re: (Score:2)
Re: (Score:3)
It's probably worth asking, "What are the inputs needed for this quantum computation and how do you get them to the quantum computer?"
Looking at the example given in the summary, it seems like the biggest issue would be marshalling accurate values for the data quickly enough would be a bigger problem than the computation per se. This may just be a tortured example chosen to be something that an average reader would relate to.
I think what might be more of interest to an outfit like Amazon is doing some kind
Re: (Score:2)
There's an even more important question, which is "which alien civilisation is going to come and bestow a computer that violates the laws of physics on us"? To paraphrase someone's comment in another discussion, all of these rhapsodies about how wonderful things will be once we have a quantum computer is like someone designing a faster than light spaceship, in crayons, leaving a big empty space where the engine would be, labelling it 'faster than light engine goes here'.
"Magic machine... err, quantum comp
The physics around quantum computing... (Score:2)
... isn't yet settled. Despite the hype no one is still sure if its possible outside of a white paper. I'm no expert in the subject either, but given that doing calculations to create information requires a finite amount of energy yet quantum computers seem to solve some intractable problems (for conventional computers) almost for free I really wonder if they're actually feasible in the real world. Eg the energy cost of error correcting the qbits heads off exponentially into the realm of fantasy in the same
Re: (Score:1)
Actually the physics is quite well settled, it's "just" the engineering that is not, i.e., will we ever be able to solve the technical challenges to turn theory into practice? All the recent evidence points to yes, but who knows. Think of fusion energy: "always 20 years in the future", although recent advances seem to go in the direction of finally breaking this vicious circle. But, as usual, haters gonna hate and will tell you that [insert novel technology here] is a scam.
Ah, and by the way, no, quantum co
Re: The physics around quantum computing... (Score:1)
50 years ago a home hobbyist could buy some cheap 7400 chips and play with logic - could build a full functioning computer if you were ambitious and dedicated enough.
Today I cannot find anything like that qbit wise - just the equivalent of a single gate to play with superposition etc without near millions of lab cryogenics etc
Re: (Score:2)
Quantum computing will most likely be first seen as a cloud service. We aren't going to be putting quantum computers in our pockets to begin with, if ever. It'll be more like an AWS Lambda function, it gets called over an API and the response given.
Re: (Score:2)
That's why room-temperature [sciencedaily.com] developments are so important.
Re: (Score:2)
We're nowhere near the Altair era of quantum computing, we're somewhere between the Jacquard loom and difference engine eras since there aren't even general-purpose quantum computers yet, just special-purpose machines to perform specific quantum functions. But you could put together a cryogenic vibration-isolated setup for under 4 digits with a total-loss LN2-based system suspended from strips of bicycle inner tube.
Re: (Score:2)
Iirc quantum computing doesn't violate the Church-Turing thesis, and thus offers no cheat to godlike powers. It can just scour solutions much faster taking advantage of functionally massive parallelism of quantum states.
Anyway, has it been shown to actually work? Every time I read a story, buried in the fine print is they can't prove the calculation wasn't done by classical computation.
Re: (Score:2)
So far, the best that's actually been managed is that a quantum computer is better at being a quantum computer than a simulator is MOST (but not all) of the time.
For all of the hype, that's it.
I'm not saying it'll never work or that it'll never do anything useful, just that any realistic comparison to the advancement of conventional computers would place quantum computing much closer to Charles Babbage and Lady Ada than to Alan Turing or Grace Hopper.
Re: (Score:2)
Indeed. There is a very real possibility QCs will not actually work because the Physics is slightly different from what we think it is. Even exceptionally slight unknown effects we have not yet observed can completely kill a moderately complex (i.e. not useful yet) QC. The other thing is, of course, that QCs scale abysmally bad, like inverse exponentially with effort or so and hence, if they ever work, may very well never hit useful sizes.
The whole thing is stupid. Accept that computing has pretty much hit
Irony (Score:2)
Make sense, they actually suck at this (Score:1)
Re: (Score:2)
Re: (Score:2)
'Tomorrow by 10am' could be accurate, OR it could just as often be three days later.
That's the quantum uncertainty part.
Alexa, order some cat food for the dead/not-dead cat, maybe.
Re: (Score:1)
So is it the (Score:1)
Traveling Salesman (Score:2)
Re: (Score:2)
Deciding which trucks, planes, drivers and pilots will go where is a traveling salesman problem on steroids.
You're going the wrong way; it is an anemic traveling salesman problem.
For example, the trucks and planes leave at certain times. Getting the package to the truck 1 second earlier rarely places it on a different delivery. It rarely speeds departure.
For traveling salesman to be hard, you need it to be a fake salesman; you need to keep everything in continuous variables, and care about even the smallest difference between routes. It is the last fractional second that is hard. Here you have a bunch of low-pass
Re: (Score:2)
For example, the trucks and planes leave at certain times.
If someone else has decreed what all the departure times are and you are powerless to change them, that does indeed simplify the problem for you.
OTOH if you are Amazon and own your own trucks and planes, you probably can (to some degree) choose when you want your planes and trucks to leave, which means that you now get to write an algorithm to determine what the best departure times should be, along with everything else. Fun!
Re: (Score:2)
My analysis already assumes they can change the times of the flights.
That entirely misses the point, though. Traveling salesman is already only hard when the goal is to find the infinitesimally better path. If you have 10,000 salesman (packages) that have to share 1 airplane, the opportunities to get a slightly-better-path than the cheaply-computed good path go way way down. And if, as here, most addresses only receive one delivery per day you won't have done any better at all.
Real-life traveling salesman i
You need to solve the hard one to solve the easy o (Score:2)
Knowing when it will get there is easy *if you know the routes all the trucks and Vans are going to take*. That's the USPS model - you know it'll take 8 days, because you know a truck drives from College Station to Houston every night at midnight.
Amazon's model is that the delivery vans go where deliveries are needed that day. They don't take the same route every day. Before you can know what you're going to put a certain package on a particular van, FIRST you must figure out what route that van is going
Might (Score:2)
Might. Also, might not. In other news, the following statement is true: (A or not A).
Oooh! a new buzzword! (Score:1)
"quantum consumer"!
Be sure to add it to your bullshit bingo card!
"quantum" ... Please comment if you know more such terms.
"crypto" (implying "blockchain")
"AI"
"cloud"
"smart"
"cyber"
I'm more of an old-school type:
"turbo"
"atom"
"iron"
"bronze" [youtube.com]
"wheel"
"stone"
--
If you take this comment too seriously, you will be punished with a bad mood.
Re: (Score:2)
Re: (Score:2)
Obligatory Dilbert:
https://dilbert.com/strip/1994... [dilbert.com]
"Bingo, sir"
Traveling salesman does little (Score:2)
Having the perfect answer can marginally reduce your time, but not a lot.
Re: (Score:2)
The other part of the problem is search. The example was, find all the products that meet a particular criteria that can be delivered by tomorrow. Part of that is travelling salesman, part of it is search.
But it's the kind of search that quantum computing is complete crap at: sorting through an enormous database. So the example problem is half something that QC likely doesn't offer any practical advantage over approximate classical algorithms, and half something that it's absolutely terrible at.
Considering
Re: (Score:2)
Nice !
No (Score:2)
There _still_ is no QC that actually works. And it looks more and more like there never will be one or at least there will not be one anytime soon. This has been seriously researched for something like 40 years now and there is _nothing_ real to show for it. That rather strongly indicates that this will go nowhere.
The other thing is that QC computations will never be "real time" (if they ever work): You have a lengthy set-up phase top create entanglement and initial state, then you have a slow computation,