Quantum computers capable of solving classically intractable problems are under construction, and intermediate-scale devices are approaching completion. Current efforts to design large-scale devices require allocating immense resources to error correction, with the majority dedicated to the production of high-fidelity ancillary states known as magic-states. Leading techniques focus on dedicating a large, contiguous region of the processor as a single magic-state distillation factory responsible for meeting the magic-state demands of applications. In this work we design and analyze a set of optimized factory architectural layouts that divide a single factory into spatially distributed factories located throughout the processor. We find that distributed factory architectures minimize the space-time volume overhead imposed by distillation. Additionally, we find that the number of distributed components in each optimal configuration is sensitive to application characteristics and underlying physical device error rates. More specifically, we find that the rate at which T-gates are demanded by an application has a significant impact on the optimal distillation architecture. We develop an optimization procedure that discovers the optimal number of factory distillation rounds and number of output magic states per factory, as well as an overall system architecture that interacts with the factories. This yields between a 10x and 20x resource reduction compared to commonly accepted single factory designs. Performance is analyzed across representative application classes such as quantum simulation and quantum chemistry.
Download