Amazon Braket – Get Started with Quantum Computing

* Decentralized as a part of early development

* Centralized for ease of early deployment

* Decentralized once it becomes simple / commodity enough that everyone can just have one

* Recentralized once it's cheaper to run them all centrally again

And then you only break back out once the thing you're doing fundamentally changes for some reason.

This is like dialing up to use the university PDP-11 in the 70s, basically. Big things are coming.

Even if he had said it, it was a very accurate statement at the time. Gordon Bell has noted that at the time he is claimed to have made that statement, it would have held essentially true for a decade. As something that would have likely been said (if said at all) in discussions around IBM's near-future business plans, or a sort of market analysis of the conditions at the time, it makes perfect sense.

Oh, and fusion reactors.

Edit:

Now that I think of it, there was a lot of variation in predictions of computing. But I would say that it's been pretty common for science fiction to describe technology 30 years out somewhat accurately, probably because it has inspired the actual tech in a self-fulfilling way.

So, in the 40s, a spaceship was envisioned able to carry only calculators and slide rules, with a radio link to a big central computer. That wasn't far off of how things developed in the 60s and 70s. But I think by the 60s and 70s, people were imagining pocket computers and tablets and such and that had a huge effect on people actually designing them when it was possible.

By no means do I intend to discourage progress! I wouldn't do the work that I do if it wasn't so difficult. I did hedge, a bit: "may never," "foreseeable future," "without major developments."

The fridge is just one major obstacle. There's a plethora of physics, engineering, and mathematical challenges out there impeding progress. Get to work!

That said, LH2 temperatures aren't really hard and can easily fit in a 2U rackmount device, providing power/classical-RF uses of type2 superconductors. Think EMI shielding, power conditioning, ~50 GHz traces that can span a full backplane without fancy signal conditioning, etc.

It is very hard to dissipate heat from a solid object into space. This is not true for our bodies on the other hand, but that is more to do with pressure - if you expose cells to the Void of space, most liquids inside would quickly expand in size and essentially boil, consuming large amounts of heat to go through the phase transition from liquid to gas, thus quickly cooling surrounding tissue. You could theoretically use this to create evaporation-based heating, but you would have to transport vast quantities of water that would quickly be used up, since there is no hope of collecting them back most likely.

The everyday person will use QC like they "use" machine learning today: from a very high level abstract viewpoint, where services they consume have a little bit of intelligence that makes interacting with them more efficient.

Software simulates what's supposed to happen while the hardware only approaches it through many repeated trials.

If you set up a circuit where you've entangled two qubits so that they should come out as the same value (00 or 11) and the configuration says they should come out with 50% chance of either, the simulation will show that. The outputs of 01 and 10 will never show up in the simulation. But in the real world, there's still a chance that you get those. You'll likely (on IBM's quantum computers) get something like 1-5% 01, 1-5% 10, 45-50% 11, 45-50% 00 (again, over thousands of runs).

If you want to see how this plays out with simulations and real quantum computers, IBM [0] has free access (constrained by credits when you want to run on real quantum computers, they reset each day).

[0] https://quantum-computing.ibm.com/

A big part of writing the supremacy paper was optimizing the simulators. We did three different styles of simulation:

1) A state vector simulator with hand rolled SIMD assembly (called qSim; not yet publicaly released). This required too much space at 53 qubits. (Well, unless you're going to use the majority of all disk space on summit, which brings its own obstacles. IBM says they can run it that way in a few days, but we'll see.)

2) Treating the quantum circuit as a tensor network and doing optimized contraction to avoid the space blowup (called qFlex https://github.com/ngnrsaa/qflex). This required too much time at 53 qubits.

3) Custom code written to run on a supercomputer instead of distributed computers.

There's only so much effort you can put in before you have to call it. I think it's more likely for an algorithmic break to save a factor of 10 than for optimization to save a factor of 10 at this point.. although someone should probably try using GPUs or FPGAs.

I also take the view that if it takes a month to produce a new optimized implementation of a classical simulator that beats the quantum hardware, then the quantum hardware is still outperforming classical hardware for that month. The theoretical bounds are important, but in the day-to-day context of a race they aren't directly relevant.

Part of the problem is that comparing a quantum computer simulation to the state of an actual quantum computer is almost inherently an apples-to-oranges thing.

We do not know of any way to compactly represent the state of a quantum computer in a classical computer; it appears to require an exponentially large state vector. So it requires an exponential number of operations just to evolve the full state vector. But it also produces the exact and full probability distribution describing the quantum computer state. Running a physical QC produces one _sample_ from that distribution, and you have to run it completely again to get another sample.

If we pretend that the simplest quantum computer state is equivalent to a coin toss, then classically simulating a fair coin would produce a distribution like {H: 0.5, T: 0.5}, while running the quantum computer would produce “H” half the time and “T” half the time. You could run the QC a bunch of times to estimate the actual probability, but there’s a fundamental difference between what the QC does and the simulation does.

There are some ways to approximately solve the classical simulation, but we haven’t found a way to compactly represent and manipulate the quantum state vector that are as fast as physics/nature seem to be able to do it in hardware. The Extended Church-Turing thesis (ECT) basically says that such an encoding and simulator should exist, and the basic premise of QC is that such an encoding and evolution process is impossible.

We don’t expect to ever conclusively prove that ECT is false (someone could always find an efficient simulator for QCs) but that doesn’t mean QCs can’t be useful in the interim, or even potentially after such a simulator is found (if it is a polynomial with a very large constant factor, for example).

The supremacy paper claimed that classically simulating their largest circuit would take 10,000 years. IBM said it would take them 2 days. The sycamore processor did it in 5 minutes. That still a big improvement, even if the classical simulator can be further optimized—and even IBM admits their simulator wouldn’t work if the quantum processor had even 1 more qubit.