My dissertation presents new computing hardware that can analyze large, complex networks faster and more efficiently. Many companies use such networks to understand social media connections, optimize e-commerce delivery strategy, predict weather, and more, but such data is difficult for today's computers to analyze because of the way computers store and move large datasets. I classify network-based algorithms into three types and show new computer processor designs for each, widening key performance bottlenecks by placing compute units inside the memory (instead of next to the memory). Together, these designs run their target applications up to 14x faster while using up to 93% less energy, enabling future computers to keep pace with the ever-rising demand for more compute power.
ACRE: Accelerating Random Forests for Explainability (First Author)
DyGraph: A Dynamic Graph Generator and Benchmark Suite (First Author)
GreenScale: Carbon Optimization for Edge Computing
Evergreen: Comprehensive Carbon Model for Performance-Emission Tradeoffs
PIM-Potential: Broadening Acceleration Reach of PIM Architectures
US Patent 11,803,311 - System and Method for Coalesced Multicast Data Transfers over Memory Interfaces
Now owned by AMD.