Toward More Energy-Efficient Datacenters with Enhanced Programmable Silicon

点击翻译按钮获取中文摘要

The ability to fabricate ever more transistors on a semiconductor chip has enabled ever more computation, transforming society with ubiquitous communications, advanced medical imaging, on-demand entertainment delivery and more. We have become accustomed to the cost and energy of computation dropping exponentially over time, enabling new applications and societal benefits that motivate investment in raising computational capabilities further. This virtuous cycle is now under threat as the energy required for a computation no longer drops sufficiently with improved manufacturing of conventional processors (CPUs). There is an alternative: configure specialized hardware into Field Programamble Gate Array (FPGA) chips instead of controlling general CPU hardware with software instructions. This allows FPGAs to use 10% or less of the energy required by a CPU for many computations. However, FPGAs face their own challenges. First, it is harder to design an application to run on an FPGA than on a CPU, and as improved manufacturing continually increases the logic capacity of FPGAs it is becoming very time-consuming to complete designs that fully utilize them. Second, increasingly computation is moving from special purpose systems such as cell phone base stations, which often employ FPGAs, to centralized data centers, which today almost exclusively use CPUs. Moving computation to data centers enables centralized administration and hardware sharing but Data centers consume tremendous amounts of power (~2% of US electricity demand). They could greatly benefit from the energy-efficiency of FPGAs, but the difficulty of FPGA application design and sharing between applications have limited their use in data centers. We will make today's high-capacity FPGAs easier to use by creating new modular software tools that let engineers easily connect smaller hardware blocks to create a full application. We will also investigate embedding a communication network in the FPGA chip to make it easier still for hardware blocks to communicate. We will create a new design style that allows applications to be interrupted and swapped in and out of an FPGA over time so it can be shared. Finally, we will scale an FPGA-accelerated simulator and optimizer for light-activated chemotherapy to data center scale, and show how our new tools, FPGA architectures and swappable design style benefit this application. This will both validate our FPGA improvements and help make a promising new cancer therapy practical. Together the innovations we propose will enable wider use of FPGAs in the data center, reducing energy, operating costs and environmental impact. Furthermore, as energy is the limiting factor in computation today, making more energy-efficient computing practical in data centers is essential for further scaling of computational capabilities, and the new applications and benefits it will enable.