Exponential please, and make it a double.
After 1,400 words, there’s the phrase everyone else put right in the headline. Why? Because quantum supremacy may sound grand, but it’s only a small part of what was accomplished, and in fact this result in particular may not last forever as an example of having reached those lofty heights. But to continue.
Google’s setup, then, was simple. Set up randomly created circuits of qubits, both in its quantum computer and in the simulator. Start simple with a few qubits doing a handful of operational cycles and compare the time it takes to produce results.
Bear in mind that the simulator is not running on a laptop next to the fridge-sized quantum computer, but on Summit – a supercomputer at Oak Ridge National Lab currently rated as the most powerful single processing system in the world, and not by a little. It has 2.4 million processing cores, a little under 3 petabytes of memory, and hits about 150 petaflops.
At these early stages, the simulator and the quantum computer happily agreed – the numbers they spat out, the probability spreads, were the same, over and over.
But as more qubits and more complexity got added to the system, the time the simulator took to produce its prediction increased. That’s to be expected, just like a bigger pachinko machine. At first the times for actually executing the calculation and simulating it may have been comparable – a matter of seconds or minutes. But those numbers soon grew hour by hour as they worked their way up to 54 qubits.
When it got to the point where it took the simulator five hours to verify the quantum computer’s result, Google changed its tack. Because more qubits isn’t the only way quantum computing gets more complex (and besides, they couldn’t add any more to their current hardware). Instead, they started performing more rounds of operations with a given circuit, which adds all kinds of complexity to the simulation for a lot of reasons that I couldn’t possibly explain.
For the quantum computer, doing another round of calculations takes a fraction of a second, and even multiplied by thousands of times to get the required number of runs to produce usable probability numbers, it only ended up taking the machine several extra seconds.
For the simulator, verifying these results took a week – a week, on the most powerful computer in the world.
At that point the team had to stop doing the actual simulator testing, since it was so time-consuming and expensive. Yet even so, no one really claimed that they had achieved “quantum supremacy.” After all, it may have taken the biggest classical computer ever created thousands of times longer, but it was still getting done.
So they cranked the dial up another couple notches. 54 qubits, doing 25 cycles, took Google’s Sycamore system 200 seconds. Extrapolating from its earlier results, the team estimated that it would take Summit 10,000 years.
What happened is what the team called double exponential increase. It turns out that adding qubits and cycles to a quantum computer adds a few microseconds or seconds every time – a linear increase. But every qubit you add to a simulated system makes that simulation exponentially more costly to run, and it’s the same story with cycles.
Imagine if you had to do whatever number of push-ups I did, squared, then squared again. If I did 1, you would do 1. If I did 2, you’d do 16. So far no problem. But by the time I get to 10, I’d be waiting for weeks while you finish your 10,000 push-ups. It’s not exactly analogous to Sycamore and Summit, since adding qubits and cycles had different and varying exponential difficulty increases, but you get the idea. At some point you can have to call it. And Google called it when the most powerful computer in the world would still be working on something when in all likelihood this planet will be a smoking ruin.
It’s worth mentioning here that this result does in a way depend on the current state of supercomputers and simulation techniques, which could very well improve. In fact IBM published a paper just before Google’s announcement suggesting that theoretically it could reduce the time necessary for the task described significantly. But it seems unlikely that they’re going to improve by multiple orders of magnitude and threaten quantum supremacy again. After all, if you add a few more qubits or cycles, it gets multiple orders of magnitude harder again. Even so, advances on the classical front are both welcome and necessary for further quantum development.
‘Sputnik didn’t do much, either’
So the quantum computer beat the classical one soundly on the most contrived, lopsided task imaginable, like pitting an apple versus an orange in a “best citrus” competition. So what?
Well, as founder of Google’s Quantum AI lab Hartmut Neven pointed out, “Sputnik didn’t do much either. It just circled the Earth and beeped.” And yet we always talk about an industry having its “Sputnik moment” – because that was when something went from theory to reality, and began the long march from reality to banality.
That seemed to be the attitude of the others on the team I talked with at Google’s quantum computing ground zero near Santa Barbara. Quantum superiority is nice, they said, but it’s what they learned in the process that mattered, by confirming that what they were doing wasn’t pointless.
Basically it’s possible that a result like theirs could be achieved whether or not quantum computing really has a future. Pointing to one of the dozens of nearly incomprehensible graphs and diagrams I was treated to that day, hardware lead and longtime quantum theorist John Martines explained one crucial result: The quantum computer wasn’t doing anything weird and unexpected.
This is very important when doing something completely new. It was entirely possible that in the process of connecting dozens of qubits and forcing them to dance to the tune of the control systems, flipping, entangling, disengaging, and so on – well, something might happen.
Maybe it would turn out that systems with more than 14 entangled qubits in the circuit produce a large amount of interference that breaks the operation. Maybe some unknown force would cause sequential qubit photons to affect one another. Maybe sequential gates of certain types would cause the qubit to decohere and break the circuit. It’s these unknown unknowns that have caused so much doubt over whether, as asked at the beginning, quantum computing really exists as anything more than a parlor trick.
Imagine if they discovered that in digital computers, if you linked too many transistors together, they all spontaneously lost their charge and went to 0. That would put a huge limitation on what a transistor-based digital computer was capable of doing. Until now, no one knew if such a limitation existed for quantum computers.
“There’s no new physics out there that will cause this to fail. That’s a big takeaway,” said Martines. “We see the same errors whether we have a simple circuit or complex one, meaning the errors are not dependent on computational complexity or entanglement – which means the complex quantum computing going on doesn’t have fragility to it because you’re doing a complex computation.”
They operated a quantum computer at complexities higher than ever before, and nothing weird happens. And based on their observations and tests, they found that there’s no reason to believe they can’t take this same scheme up to, say, a thousand qubits and even greater complexity.
That is the true accomplishment of the work the research team did. They found out, in the process of achieving the rather overhyped milestone of quantum superiority, that quantum computers are something that can continue to get better and to achieve more than simply an interesting experimental results.
This was by no means a given – like everything else in the world, quantum or classical, it’s all theoretical until you test it.
It means that sometime soonish, though no one can really say when, quantum computers will be something people will use to accomplish real tasks. From here on out, it’s a matter of getting better, not proving the possibility; of writing code, not theorizing whether code can be executed.
It’s going from Feynman’s proposal that a quantum computer will be needed to using a quantum computer for whatever you need it for. It’s the “hello world” moment for quantum computing.
Feynman, by the way, would probably not be surprised. He knew he was right.
Google’s paper describing their work was published in the journal Nature. You can read it here.
Source: TechCrunchRelated posts: