Efficiency; Redox flow battery; Computational modeling; Power; Microfluidics; 3D Integration; Density; Computer ; Fluid dynamics; In situ analytics; Electrocatalysis; Chaotic mixing; Selective membranes; Heat and mass transfer; Microfabrication
Marschewski Julian, Jung Stefan, Ruch Patrick, Prasad Nishant, Mazzotti Sergio, Michel Bruno, Poulikakos Dimos (2015), Mixing with herringbone-inspired microstructures: overcoming the diffusion limit in co-laminar microfluidic devices, in LAB ON A CHIP
, 15(8), 1923-1933.
Sabry Mohamed M., Sridhar Arvind R., Atienza David, Ruch Patrick, Michel Bruno (2014), Integrated microfluidic power generation and cooling for bright silicon MPSoCs, in Design, Automation and Test in Europe Conference and Exhibition (DATE), 2014
Sridhar Arvind, Sabry Mohamed M., Ruch P., Atienza D., Michel B., PowerCool: Simulation of Integrated Microfluidic Power Generation in Bright Silicon MPSoCs, in 2014 IEEE/ACM International Conference on Computer-Aided Design,
We address here the integration density of future high-performance computers with an approach inspired by packaging and architectural principles of the human brain: a dense 3D architecture for interconnects, fluid cooling, and power delivery through redox species transported in the coolant with low power requirement for pumping. Vertical integration improves memory proximity and bandwidth, but current power delivery and cooling solutions do not allow integration of multiple layers with dense logic elements. Interlayer liquid-cooled 3D chip stacks solve the cooling bottleneck but are still limited by power delivery and communication thresholds. We propose a fundamental concept, redox flow electrochemistry for power delivery and cooling (REPCOOL), which eliminates the conventional electrical power supply network, thereby reducing conversion and transport losses, liberating valuable space for communication, and allowing scaling to systems beyond exascale device count and performance. This novel bionic concept is similar to the integration principle in the human brain with fluid-based power delivery and promises to improve system efficiency by several orders of magnitude. We aim at addressing this challenging research through the following synergistic research packages:(1)The first research component will establish a property catalog required for on-chip integration of REPCOOL modules. Relying on existing microfabrication and packaging expertise, microfluidic cells will be fabricated for component testing by detailed spatiotemporally resolved electrochemical and thermal characterization. At the end, all sub-project components will be integrated into a joint demonstrator with ~10 W/cm2 electrochemical power density and <15 K temperature rise of the coolant using electrical, thermal and fluidic packaging established by this first sub-project. (2)The second research component focuses on advanced electrode and membrane materials for electrochemical conversion to substantially enhance reaction kinetics and ionic selectivity. The main tasks involved are fabrication and electrochemical screening of electrode and membrane materials for REPCOOL systems and their characterization in terms of structure and electrochemical performance.(3)The third component focuses on markedly improving heat and mass transport in the microfluidic REPCOOL systems. This effort encompasses flow visualization and characterization with electrochemical modeling and implementation of transport enhancement schemes.(4)The fourth research component will focus on enhancing the scientific understanding of interactions between redox couples and surface chemistry using ab initio molecular dynamics. This study encompasses a predictive computational study of electrolyte speciation and electrode surface chemistry and their influence on reaction kinetics to guide experimental system development. All these components aim at developing a science base for combined power delivery and cooling for the next generation microprocessors, and clearly require a collaborative, multidisciplinary research effort. The power densities required to feed chip stacks are only possible when mass transport is maximized with minimal convective resistance (3), interfacial redox reactions are catalytically favored and electrode/membrane design is optimized (2), which can only be successfully solved in a short time with the benefit of predictive computational modeling (4) and dedicated experimental flow cell trials (1). A final joint consolidation of the results then allows reaching the power densities necessary to impact ultra-dense, high efficiency chip stacks. The proposed effort will contribute to the fundamental understanding of future dense compute architectures that are also needed for neuromorphic hardware and will help advance the efficiency of computing systems (today ~10^9 ops/J) toward the efficiency of biologic brains (~10^14 ops/J).