Riverlane summarises a new paper - ‘Practical Quantum Computing: the value of local computation’ - just published:
These devices will rely on classical components to supplement calculations. Optimising the delicate interplay between central processing units (CPUs) and quantum processing units (QPUs) will identify complex operational challenges that must be overcome to truly capture the benefits of quantum computing. Developing a hardware-conscious solution to these problems is one of the most pressing issues in near-term quantum computing.
Current quantum computers employ a `black-box’ model: quantum programs are written on a CPU and forwarded to the QPU to be implemented blindly. This model does not allow more intricate hardware-conscious programming to take place. Here we present three broad areas where this approach causes performance bottlenecks – an issue not often discussed within the quantum computing community:
- Bandwith issues related to limited communication capacity between CPU and QPU;
- Latency bottlenecks due to the round-trip time delay;
- Build-up of qubit error rates due to short qubit coherence times.
In each case we present an algorithm to emphasise the prohibitive impact these bottlenecks can have on a computation: randomised benchmarking, adaptive quantum chemistry algorithms, and quantum error correction respectively. Our examples demonstrate that particular algorithms need a large instruction bandwidth, or require certain calculations to be performed very frequently and at regular intervals, to determine the next step in the algorithm. We show that this leads to poor performance within the current computational paradigm for quantum computers.
We propose a mitigating strategy that takes advantage of the classical hardware that is situated in close proximity to the QPU. In most quantum computers this hardware is comprised of field programmable gate arrays (FPGAs). The types of processes that are well suited to being implemented by this local classical hardware are ones that are relatively simple and occur frequently. We identify the elements of randomised benchmarking, adaptive quantum chemistry algorithms, and error correction that have this characteristic, and by allowing the local computation to handle these algorithmic components, we show that the restrictive problems with the black-box model can be overcome.