The Long and Winding Road to Energy Efficiency Using Customized Computer Architectures

Abstract: Customized computer architectures are specially designed processors and accelerators that match the hardware capabilities with the underlying computation, memory, and control behavior of a target application domain. Specialization enables traditionally conflicting goals to be achieved simultaneously in comparison to general-purpose solutions: higher performance and reduced energy consumption, cost, and complexity. In this talk, I will highlight two important findings discovered on the road to creating customized hardware. First, software should not be considered a static input to hardware customization, but rather an often-untapped customization opportunity to push efficiency higher. In the extended reality (XR) domain, customization of the XR software stack was performed with SlimSLAM, a domain-specific runtime scheduler that detects and adjusts over-provisioned parameters in real-time to eliminate unnecessary computation. SlimSLAM outperforms other adaptive approaches by an average of 2.3x with iso-accuracy. Second, programmability is critical in hardware customization to ensure customizations are usable as applications go through inevitable changes and new ones come to the forefront. For sparse data access, a programmable data access accelerator, DX100, was designed to offload bulk indirect memory accesses and associated address calculation operations. DX100 reorders, interleaves, and coalesces memory requests to improve memory row-buffer hit rate and bandwidth utilization and achieve performance improvements of 2.6x over a traditional processor.

Bio: Scott Mahlke is the Claude E. Shannon Professor of Engineering Sciences in the EECS Department at the University of Michigan. He leads the Compilers Creating Custom Processors research group that focuses on hardware/software technologies for scaling performance, energy efficiency, and cost of computing systems through specialization of the hardware down to the software it runs. Mahlke has won numerous awards including the 2022 IEEE B. Ramakrishna Rau Award, and is a Fellow of the IEEE and ACM.