Heterogeneous Compute Architectures for Deep Learning in the Cloud
Ken O'Brien
Xilinx Research, IE
ABSTRACT
The lab-on-a-chip (LoC) concept was originally intended to introduce the irresistible and steady increasing performance of integrated circuit (IC) technology in microfluidics to automate and to parallelize the processes of complete chemical and biochemical laboratories on a single tiny microchip - a ground-breaking concept especially for life science and analytics. However, LoCs have not yet evolved revolutionary impact in microfluidics as known from microelectronics in information technology. The indisputable reason for that is the lack of scalability of the existing microfluidic platforms, especially the predominating microelectromechanical systems. The perspective solution of the basic challenge of scalability is hoped for the logic microfluidics, which is already named as “second breath of microfluidics” with revolutionary potential. The basic idea of this concept is to drastically reduce the effort for the off-chip control by integration of parts of the chip program by valve-based circuits. Currently there are two major approaches for. The micropneumatic logic microfluidics is technologically based on the multi-layer soft lithography and requires pressure-switchable valves, which derivate from pneumatic membrane valves. The second approach is the concept of chemo-fluidic logic microfluidics, the so-called chemofluidics. It is based on components with decision-making functionality for chemical information, the chemo-fluidic switches, chemostats and transistors and allows to fabricate integrated circuits by patterning of active material layers as known from microelectronics. Because of the chemical feedback the chip control is placed directly on the chip and exclusively controlled by chemical and other events. Since the components of the chip take their energy needed directly from the fluidic chemicals the chemo-fluidic logic chips are fully autonomous micro-systems processing their chip programs automatically and self-powered. A first set of hardware instructions including combinatorial logic gates, RS-flipflops and oscillators principally enabling to realise each more complex function is demonstrated. However, the concept of chemofluidics is still in its infancy. Many challenges on the levels of technology, design and system still await resolution.
