Texas A&M Researchers Develop ‘Skia’ to Boost Data Center Processor Efficiency

Researchers at Texas A&M University, in collaboration with Intel, AheadComputing, and Princeton University, have introduced a novel processor technique called ‘Skia’ designed to enhance the efficiency and performance of computer processors, particularly in data centers. Modern data centers process vast amounts of information, and the ability of processors to predict and prepare instructions is a significant bottleneck, often slowing down data flow and increasing energy consumption. Skia addresses this by improving the prediction of future instructions, thereby increasing throughput-the number of instructions executed per unit of time. The technique specifically targets inefficiencies in Fetch Directed Instruction Prefetching (FDIP), a system that relies on predicting and prefetching instructions. As workloads become more complex, the Branch Target Buffer (BTB), which tracks instructions, can fail, resulting in ‘shadow branches’-unused instructions that remain undecoded and waste cache space. Skia identifies and decodes these shadow branches, storing them in a dedicated buffer for future use, which can be accessed alongside the BTB. This approach allows for more accurate instruction prediction with minimal additional hardware, nearly doubling performance improvements compared to simply increasing storage. The research suggests that even a 10% efficiency gain could significantly reduce the number of data centers needed, saving millions in costs and reducing energy consumption. The findings were presented at a leading computer architecture conference and highlight the potential for Skia to make data centers more sustainable and responsive.