An Optical Interconnect for Modular Quantum Computers

 

Here's our latest paper! Our biggest effort of the year, or more correctly, two years of work supported by 20 years of learning and doing. There are "only" a dozen authors on this, but there are about 150 people in the Shota Nagayama Moonshot project doing supporting and related work.

This paper combines experimental demonstration of our network prototype (the portions in pink in the figure), insight into network topology (both the larger diagram here and some more complex diagrams in the paper), and an operation scheduling and performance estimation tool (not shown) to help us evaluate where we are and where we need to be.

Slightly paraphrasing the paper,

The current focus of modular quantum system research is the link and connection layers. This is a necessary first step towards building more complex systems. However, in order to bring us closer to scalable and robust distributed quantum systems, there is a pressing need for deeper understanding of higher-layer concepts such as switching network architectures, and their impact on the performance when executing large-scale quantum computation. The time for such studies has come.

We have dubbed our interconnect "Q-Fly", following on from butterfly, k-ary n-fly, Dragonfly, Dragonfly+, Slimfly, Polarfly, etc. Designing this was a deeper task than the final diagram reveals, and involves balancing the number of Bell state analyzers and the number of inter-group connections to accommodate both expected traffic patterns and link fault tolerance while maintaining minimal-length paths.

Of course one of the questions is how this Dragonfly-inspired network compares to a fat tree. In the technology we are working with, node-to-node connections are created by routing one photon from each node to a shared Bell State Analyzer (BSA).  It turns out that effective placement of the BSAs in the network is critical, and that doing a good job of that in a fat tree is harder than for our Q-Fly. The physical construction to allow some traffic to avoid going through the root of the tree is complex and involves inefficient pools of BSAs distributed awkwardly in the network. The Q-Fly design is much cleaner, and minimizes the number of switches each photon passes through even in large networks. We have to worry about every switch, since we can't afford to lose photons any more often than absolutely necessary. This principle led us to the Q-Fly mantra,

Every decibel matters.

Of course I'm riding high on finishing (well, almost -- we still have to get the paper through peer review) this paper, but as of right now I'd rate this the most important paper of my career to date.

Below is one of the early sketches on the whiteboard.