The Empire Strikes Back: Quantum Volume 64 now also at IBM

To those of us who are familiar with IBM’s devices, the new “Falcon” processors look markedly different. The reason is that the qubit layout is tailored for another quantum error correcting code (arXiv:1907.09528 — not the standard surface code). The orange qubits were used in the experiment; the qubits linked to an R can be read out simultaneously. (Figure from: Demonstration of quantum volume 64 on a superconducting quantum computing system (IBM), arXiv:2008.08571.)

Quantum volume measures the computational power of near-term digital (read: qubits) gate-based quantum computing systems. The following approximation makes clear what it means: If on n qubits you can run a quantum circuit consisting of random 2-qubit gates (uniformly in SU(4)) of depth n (assuming all-to-all connectivity), and the measurement results make sense, then the quantum volume is 2ⁿ.

IBM introduced QV in 2018 (arXiv:1811.12926). Their superconducting qubit based systems from 2017 achieved QV 4, and IBM has the declared goal to double the quantum volume every year: QV 8 in 2018, QV 16 in 2019, QV 32 in January 2020 (see this blog post) — check. In June, though, Honeywell, which pursues trapped-ion based quantum computing systems, has claimed QV 64 for their system. (Honeywell was criticized for not publishing sufficient specs of their system and I don’t know if that has changed in the meantime.) IBM must have been pissed about being beaten in the game they invented: Yesterday they published an article in which they, too, claim to have squeezed a Quantum Volume of 64 into one of their systems (arXiv:2008.08571).

New quantum computing records are everywhere these days. What makes IBM’s announcement special is that it is a beautiful illustration of the crucial role of software in near-term quantum computing.

Trapped-ion qubit systems have two advantages which help them beat superconducting qubit systems in the QV game. (1) As long as you only have a small number of qubits (less than 30, say), ion traps allow for all-to-all connectivity between the qubits. Superconducting systems, on the other hand, have resonators between pairs of qubits, and no other qubit pair can participate in a hardware gate: They must be realized in software. (2) The ions which hold the qubits in ion-trap devices are error free when idle, while superconducting qubits leak information in various ways within microseconds.

The above mentioned paper by IBM describes the software components which IBM has improved to make QV 64 happen against Honeywell’s architectural advantage. I’m going through them quickly.

(1) Improved circuit mapping and optimization (“compilation”) allows to reduce the total number of elementary quantum logic gates that are being run, hence reducing the effect of quantum noise. That helps to combat the ion-trap advantage #1 above. (2) Error mitigation for qubits that are idle to reduce the noise-advantage of ions as qubits. (3) Improved quantum control. As quantum gates are ultimately run by sending analog microwave pulses onto the chip, realizing something digital and clunky as a quantum logic gates requires some careful fiddling with these pulses. That is the job of the quantum control, a software which aims to send the optimal pulses for the desired gate, while at the same time keeping the noise down.

Und die Moral von der Geschicht…

Software plays a vital role. (We already knew that, but, as always in teaching, redundancy helps.) To participate in the quantum computing revolution we don’t need to solder our own quantum computer in the basement. There’s a whole new stack of software waiting to be researched out. Just think of all the software components in a classical computer that you take for granted. All of that has to be invented, tried out, improved, and perfected over the next 15 years. I recommend the paper to anyone who wants to get an idea of some of the lower levels of the software stack.