Design and management of three-dimensional multi-processor systems-on-chip with integrated flow cell arrays

Three-Dimensional Multi-Processor Systems-on-Chip (3D MPSoCs) are promising solutions for highly intensive Artificial Intelligence (AI) and Big Data applications. They combine remarkably dense computation capabilities and massive communication bandwidths. However, due to their high density, 3D MPSoCs present heat dissipation, and power delivery challenges. Flow Cell Arrays (FCAs) have the potential to solve both issues. They consist of micro-channels etched in the Silicon substrate of stacked dies, filled with an electrolytic flow that absorbs the generated heat and produces power through electrochemical reactions. Hence, FCAs enable concurrent on-chip liquid cooling and electrochemical power generation in 3D MPSoCs.

In this context, this thesis focuses on the system-level integration of FCAs as an effective cooling solution and power source for next-generation 3D MPSoCs. First, a comprehensive framework is proposed to model 3D MPSoCs with FCAs in fine grain and analyze their thermal and power performance. Simulations demonstrate temperature reductions of up to 40°C and voltage drop reductions of up to 53% when using FCAs, even for 3D MPSoCs fabricated using deeply-scaled CMOS technologies, which are characterized by extremely high power densities. Next, the thesis introduces design-time techniques to enhance the performance of FCAs. It advocates the use of Switched Capacitor (SC) converters as an interface between the channels and the power delivery grid. SCs enable FCAs to operate at their optimal voltage to maximize power generation, resulting in gains of up to 2x compared to direct connectivity to the power grid. Then, considering FCA's interdependent thermal and power generation capacities, several design configurations are explored, highlighting trade-offs and revealing opportunities to increase the power budget of 3D MPSoCs without violating design constraints. Finally, the thesis illustrates a novel strategy to manage the cooling and power generation capabilities of FCAs at run-time. Such an optimization strategy boosts the operating frequencies of dies, increasing the computing performance up to 24% while reducing coolant flow requirements.