GPU systems through tensor contractions and power-aware benchmarking
Building GPU systems through tensor contractions and power-aware benchmarking, with a focus on hardware understanding, memory movement, and energy efficiency.
About
I build around tensor contractions as a way to understand how computation maps onto hardware.
tensorcontract is the foundation for exploring tensor algebra, contraction order, data layout, memory traffic, and GPU execution.
gpu-watt-watcher adds power-awareness and helps measure how much energy GPU workloads use across compute, memory, and AI-style tasks.
The goal is to connect performance, efficiency, and hardware intuition, not just raw throughput.
We study GPU computation from both a structural and energetic perspective. Tensor contractions are used as a lens for understanding how algebraic structure maps onto GPU hardware, while power measurement helps evaluate the energy cost of those computations in practice.
Projects
Open-source tools for understanding GPU structure and energy.
tensorcontract
Foundation for exploring tensor algebra, contraction order, data layout, memory traffic, and GPU execution.
gpu-watt-watcher
Power-aware benchmarking for compute, memory, and AI-style GPU workloads. Measures energy, performance, and efficiency trade-offs in real time.
Contact
Open to collaborations on GPU systems, energy-efficient machine learning, and tensor-network methods.