Superconducting Asynchronous Logic for Ultra-low Power High Performance Computing
High performance computing is bottlenecked by increasing power demands and memory bandwidth, while superconducting electronics are bounded in circuit complexity due to a limit on the number of switching devices on a single chip. This thesis proposes a modular, asynchronous superconducting computing...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/142825 https://orcid.org/0000-0002-7435-3840 |
| Summary: | High performance computing is bottlenecked by increasing power demands and memory bandwidth, while superconducting electronics are bounded in circuit complexity due to a limit on the number of switching devices on a single chip. This thesis proposes a modular, asynchronous superconducting computing framework which aims to solve both of these problems. A discrete set of logic gates are proposed and implemented using Adiabatic Quantum Flux Parametron (AQFP) logic. AQFP logic devices can achieve picosecond gate delays with zeptojoule (10−21 J) switching energy, just bordering the theoretical Landauer limit for computing energy demands, by adiabatically switching the location of a single flux quanta in a double-well potential. The heart of the project lies in the modular architecture design that realigns hardware layout with software dataflow to allow for scalable, distributed computing systems from basic circuit building blocks. Projecting the simple circuit design performance to large-scale high performance computing systems, Super-DICE aims to achieve a 103 order of magnitude improvement in power consumption, while still accounting for the cryogenic cooling overhead of the superconducting electronics. Beyond the dramatic power performance improvement with this logic technology and architecture, it also allows for designers to rapidly prototype hardware computing optimizations without needing to go through the expensive and time consuming process of fully custom ASIC design.
In this thesis, I review the device physics of the Quantum Flux Parametron and present a set of basic AQFP combinatorial logic gates. I then propose a circuit design for asynchronous token buffering between these modular gates and describe how they can be assembled as digital materials to create scalable, complex 3D computing structures. I simulate the proposed circuit designs in SPICE and project performance of a potential superconducting supercomputer using this framework. Motivated by the energy efficiency of superconducting electronics, the heart of this thesis radically proposes to redefine traditional processor architecture by discretizing large-scale system integration into a heterogeneous set of building blocks which blur the line between hardware and software with a reconfigurable, asynchronous spatial computing system. |
|---|