The research in the Vertical group spans VLSI technology to software and application analysis. Such an integrated approach is necessary to design future high performance systems. Our research focuses on future systems targeting a 2020 time-frame considering three broad directions:
Driving much of our work is detailed technology models and analytical tools.
Devices are becoming increasingly brittle, highly varying in their properties, and error-prone, leading to a fundamentally unpredictable hardware substrate. The model of hardware being correct all the time, on all regions of chip, and forever, becomes prohibitively expensive to maintain. Emerging new classes of applications are increasingly relying on probabilistic methods. They have an inherent tolerance for uncertainty, do not require hardware to be correct all the time, and this provides an opportunity to creatively utilize hardware. We are investigating ways to expose hardware reliability at the device level through layers of the system stack up to the application. We are building system software, architecture, and microarchitectural mechanisms that can scale to future technologies.
We are revisiting some of the fundamental principles of processor microarchitecture design in the light of energy and reliability becoming primary constraints. In the Idempotence project, we observe that the mathematical property of idempotence which allows an operation to be performed multiple times producing the same results provides a powerful and elegant way to radically simplify and eliminate many hardware structures.
The DySER project complements the Idempotence project in looking at ways to complement the core and specialize it various ways based on the underlying workload. Our DySER prototype RTL design and compiler should be available for download soon. The motivation for DySER was our early work on building a specialized multi-core architecture for ray tracing.
We build detailed analytical models and workload analysis to understand the constraints and limitations of existing systems and to project performance of future systems we explore.