Energy Proportional Future Chip-to-chip Computing Interconnect Designs

Multichiplet systems can be utilised to build a flexible, reconfigurable board-level computing platform with enhanced resource consumption and energy efficiency. Optical interconnects feature high IO capacity, distance-independent energy consumption, and low-delay connection. Optical linking fabric allows comprehensive customization of system architecture and resource allocation, which is not possible with electrical approaches. However, current optical technologies create static power overheads that reduce performance and dynamic power. This thesis offers and assesses unique architectural approaches for power-efficient multichiplet board-level optical connection usage. First, we recommend "unorthodox" UMA multi-chiplet architectures for on-board computers. We show it is only viable with optical interconnects, and simulations show it can improve execution speed and energy efficiency for a variety of workloads. We make our own simulation systems to evaluate the design trade-offs and rules for the network utilization methods are theoretically derived while the energy-delay products are investigated under various launch conditions. Then, we offer 'RENU,' a unique on-board optical interconnect control technique assessed with wavelength and spatial optical switching fabric. RENU maximises utilisation by rearranging interconnects depending on application network traces. Energy consumption is reduced, especially in low-use paradigms, compared to state-of-the-art. We then propose 'Min-ORUM', an online utilisation maximisation technique that reduces energy consumption for a variety of applications. Finally, we present 'SO-RA', a novel laser distribution control system using sophisticated SOA-based switch matrics. It reduces static power losses by reconfiguring in nanosecond time.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2023

Author dc:creator

• Sharma, Arastu

Advisors dc:contributor.advisor

• Penty, Richard
• White, Ian

• Machine learning
• Mutlichiplet
• Chiplet
• Chip-to-chip
• Interconnect
• Data-centre
• Optical Networks
• Optical Interconnect
• optical architecture
• multi-chiplet architecture
• VCSEL
• On-board computing
• on-board
• efficient computing
• energy-proportional
• energy proportional
• board-level
• reconfigurable architecture
• reconfigurable
• optical switch
• SOA switch
• silicon photonics
• silicon
• exascale computing
• WDM
• wavelength division multiplexing
• network-on-chip
• photonic switch

Licence

• https://www.repository.cam.ac.uk/bitstreams/2eb9e59e-725c-4e54-99de-65eaa27e969d/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.100567

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/355662