In the wee hours of Wednesday morning, IBM gave an unwary world its first publicly accessible quantum computer. You might be worried that you can tear up your passwords and throw away your encryption, for all is now lost. However, it's probably a bit early to call time on the world as we know it. You see, the whole computer is just five bits.
This might sound like some kind of publicity stunt; maybe it's IBM's way of clawing some attention back from D-Wave's quantum computing efforts. But a careful look shows that serious science undergirds the announcement.
The IBM system is, on a very superficial level, similar to D-Wave's. Both systems use superconducting quantum interference devices as qubits (quantum bits). But the similarity ends there. As IBM emphasizes, its quantum computer is a universal quantum computer—which D-Wave's is not.
Another big difference: IBM can address and measure the state of each qubit individually. The company can measure (and has) all the critical features of its device. If you want to know how long a qubit retains its state, IBM can tell you. IBM even shows that addressing multiple qubits in a random way doesn't affect the state of the others too badly. Big Blue is really building its quantum computer from the foundation up, while still ensuring that the engineering fits real-world requirements.
And we know a fair bit about the hardware. In 2015, IBM released a detailed schematic of how the circuit is put together and how it is linked to the outside world. The schematic probably isn't detailed enough for me to build the circuit in my local cleanroom, but I bet my friends down the hall who work in the field can. Releasing this sort of detail is what we in the industry call "doing science."
I was lucky enough to get hold of a draft of the paper about the device. The key focus of the hardware at the moment is not on computing (which is pointless with five qubits) but on ensuring that the device computes reliably thanks to good error correction.
Let me get my red pen
Modern memory and communications systems would not function without error correction. The essential idea is to build some redundancy into the data so that you can always tell when data is corrupted and sometimes correct the problems without having to ask for information to be resent.
A common scheme in classical communications is to take a group of bits, say four, and perform a series of mathematical operations (typically exclusive-or operations) on them to produce a single bit of extra data (note that I'm not outlining an exact scheme here, so in detail the exact numbers change slightly). All five bits are sent to the receiver, which performs exactly the same operations.
One can always tell if there is an error in any of the four data bits based on whether the operations produce the same value for the fifth bit. Because of the way the sequence of operations is structured, somewhere between one and two errors can be corrected based on the outcome of the operations. Basically, at a cost of about a quarter of your communication channel's capacity, you can ensure that almost no uncorrectable errors are received.
There are, of course, a whole bunch of implementation details beyond the math that you have to consider to make this work. But it does work. When I briefly worked in the microwave link industry, with error correction turned off, we expected one bit in a billion to be wrong; with error correction on, it was expected to be better than one bit per trillion.
However, this is for classical communications and computation. The quantum world, as usual, is a whole different story.
For quantum error correction, you need a quantum red pen
So, in addition to a bit-flip error, you can also have a phase-flip error, which is an error in the relationship between two bits, rather than an error in the value of one particular bit.
Quantum computing has a problem with error correction—a far worse one than classical computing faces. Let's put it in perspective. One option for a qubit is a superconducting quantum interference device where the typical energy difference between a one and a zero is on the order of 10-24 joules. For a classical system, we can choose whatever energy difference we want by setting thresholds for voltages and/or currents, so we set it to something convenient, usually something on the order of a volt.
To come close to having the same 10-24J difference between a one and zero, any chip would have to operate with a gate voltage of ten microvolts. So, at minimum, classical and quantum computers operate at energy scales that are different from one another by about a factor of 10,000.
Making matters worse, a qubit state is not a one or a zero but is a probability to produce a one or a zero on measurement, and this probability evolves in time. I can take two qubits that are set identically and then measure them some time later. After repeating this many times, I should find that the probability distributions are the same. That's the theory; in practice, they experience slightly different environments, so at some point later in time, one will have slightly changed relative to the other in an unpredictable way.
So, in addition to a bit-flip error, you can also have a phase-flip error, which is an error in the relationship between two bits rather than an error in the value of one particular bit.
And this makes error correction for a quantum circuit very difficult—and very, very necessary. In classical computers, the first processors could be implemented with minimal (or no) error correction. Quantum computing will require sophisticated error correction from the get go. It is not just a case of flipping a bit; you have to know how the different qubits evolve differently in time and how to correct for that before you measure them.
Let me give you an idea of how difficult this is. A typical qubit might have a lifetime of about 50 microseconds and a coherence time of 20 microseconds. What does this mean? It means that once your qubits are set, you have to apply error correction within the first 20 microseconds and complete a step of your calculation within 50 microseconds. That doesn't seem too bad, right?
But there's always a "but." The operations to manipulate a qubit—either to perform a logical operation or to correct an error—involve microwave pulses that have a certain amount of energy. They can either be short, sharp pulses or long, slow pulses, as long as the area of the pulse remains the same. Unfortunately, short, sharp pulses cause problems, so typical pulses are 50 to 60 nanoseconds long. Given that you need to have everything done within 50 microseconds, this means you have a total of 1,000 operations for calculation and error correction. This makes it an extra-tough problem.
Making bad neighbors clean up after themselves
The IBM researchers have solved this extra-tough problem. To simplify it considerably, they have a cluster of four data qubits that are coupled to a single qubit, called the syndrome bit. The state of the syndrome qubit depends on the states of all the other qubits. This connectivity is used to determine the rate at which qubits induce errors in one another, which is then used to correct errors before the computation is complete.
The system took me some time to puzzle out, but essentially it works like this. The qubits are electromagnetic waves oscillating back and forth in an electronic device. A small amount of the wave is coupled out and mixed with the wave from the neighboring qubit. If the two qubits are in the same state, then their electromagnetic waves combine in phase and provide a strong signal to the syndrome qubit. But if the two qubits are out of phase, they cancel and provide no signal to the syndrome qubit. Likewise, the second pair of qubits do the same, providing their own combined signal to the syndrome qubit. This is called interference.
The interference between the four data qubits controls the state of the syndrome bit. This can then be used to correct single errors that crop up in the four data qubits (e.g., a flip of one qubit). Essentially, the state of the syndrome qubit controls where the signal that fixes the error is sent.
To demonstrate how effective this error correction scheme is, the researchers focused on errors introduced through cross-talk. Cross-talk occurs when a qubit that is not involved in the current operation causes the value of a qubit that is involved to flip. One example of this is when the value of one of the four qubits is used to conditionally flip the value of a second. The two remaining qubits have a set state but are not involved in this particular operation. However, because all the qubits are coupled, the two bystanders can introduce an error by causing either the control or the data bit to flip states unexpectedly.
The critical thing about IBM's system is that it can correct the error during the computation. That is: we don't read out the qubit values and then attempt error correction. Instead, we anticipate that a bystander qubit will generate a wave that interferes and induces an error. This can be canceled out by understanding how the error will be introduced and undoing it before it causes a problem.
Simply put: the researchers don't know the state of any of the qubits, but they do know how to flip states, and they know the rate at which qubits drive each other to flip. They use that knowledge for error correction.
The operation takes a certain amount of time. Over this time, the state of the bystander qubit interferes via an electromagnetic wave with a certain phase and amplitude. The researchers wait a bit, then flip the bystander qubit. This keeps its amplitude the same but inverts the phase, causing constructive interference to become destructive interference. After waiting for the same amount of time, the net effect of the bystander qubit is exactly nothing, as the inverted phase undoes everything that the qubit initially did to its neighbors.
After that, the researchers flip the bystander back to its original state so that it can also be used without error. By carefully sequencing the operational pulses and these error-correcting pulses, researchers can reduce the probability of errors significantly.
How do I get to play with this?
IBM didn't provide a location for experiencing this fancy new device. But a quick Internet search revealed that you can get access here (registration required). Apparently, IBM really wants people who actually have a serious need for quantum computing to sign up, so I haven't actually tried it myself.
In any case, five qubits is still too few to do anything useful. But, if you think you might have a good reason to use quantum computing in the future, I suggest you register and play around, because IBM has a very aggressive timetable for scaling: it expects to hit between 50 and 100 qubits within the next decade. At 50 qubits, IBM will be able to do useful stuff. That means useful qubit numbers should be coming within five years and toys that do neat tricks a couple of years later.
If you want to be ready, it might be best to spend some time understanding the system now.
reader comments
79