580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can CPU and GPU be efficiently combined to optimize heterogeneous computing performance?
This theme investigates strategies for workload partitioning, memory management, communication, and synchronization to maximize overall system performance when CPUs and GPUs operate cooperatively. Such heterogeneous computing leverages fundamentally different architectures—the latency-optimized CPU and throughput-oriented GPU—to accelerate data-intensive and parallel applications. Proper optimization must address load balancing, data access latency, and minimizing communication overhead to realize the full potential of hybrid systems.
Key finding: This chapter identifies and analyzes several critical optimization strategies for hybrid CPU-GPU systems: partitioning and load balancing computations between CPU and GPU; hierarchical memory usage to hide data access latency... Read more
Key finding: This study develops a hybrid analytical performance modeling approach to decide at runtime whether a computing kernel should execute on CPU or GPU within an OpenMP environment. Key contributions include modeling GPU memory... Read more
Key finding: This work characterizes the impact of GPU grid geometry (block and thread configuration) on kernel performance in OpenMP offloading contexts, demonstrating substantial performance improvements—up to 7x speedup—over naive... Read more
Key finding: AEminiumGPU framework facilitates writing data-parallel programs that transparently execute on CPUs or GPUs by compiling Java Map-Reduce style programs into hybrid executables. A notable contribution is the use of machine... Read more
2. What programming models and architectural features enable effective GPU task scheduling for dynamic and irregular workloads?
Traditional GPU programming predominantly targets static, data-parallel workloads, but many advanced applications have dynamic, irregular, or recursive parallelism patterns that challenge existing shading or compute language models. Research in this theme is focused on developing programming abstractions and scheduling models that support multiple instruction streams, dynamic work creation, fine-grained load balancing, data locality preservation, and varying parallelism granularities to better map irregular tasks onto massively parallel GPU architectures.
Key finding: Whippletree introduces a task-based programming model for GPUs addressing four critical requirements for dynamic irregular workloads: multiprogramming (MIMD), dynamic work generation, preserving data locality via shared... Read more
3. How can GPUs be leveraged and optimized for domain-specific high-performance computing applications?
This research area explores the implementation and acceleration of computationally demanding scientific and engineering problems using GPUs. It involves algorithm redesign, architecture-tailored numerical methods, and efficient programming to exploit GPUs’ massive parallelism and memory hierarchies. Across domains such as thermal simulations, fluid dynamics, radar signal processing, and bioinformatics, GPU computing enables orders-of-magnitude speedups, real-time analysis capability, and improved simulation fidelity, which are vital for engineering design, environmental monitoring, and biomedical research.
Key finding: This study presents a CUDA-C based framework specifically optimized for simulating conjugate heat transfer in squared heated cavities. By leveraging tailored GPU parallelization techniques, the framework achieves up to 99.7%... Read more
Key finding: This paper introduces a large-eddy simulation (LES) solver designed from scratch for modern NVIDIA GPUs with Ampere architecture to accelerate aerodynamic analysis of electric vertical takeoff and landing (eVTOL) aircraft.... Read more
Key finding: This work implements and optimizes Incoherent Scatter Radar plasma line spectral analysis algorithms on GPUs using NVIDIA CUDA. By mapping computationally expensive spectral binning, FFT, and parameter estimation operations... Read more
Key finding: The paper demonstrates a heterogeneous CPU+GPU implementation of Gaussian Process (GP) models for emulating computationally expensive simulators. By leveraging GPU parallelism for key linear algebraic operations (matrix... Read more
Key finding: This research introduces an implementation called LightKernel that exploits the cluster-based architecture of modern NVIDIA GPUs to improve predictability and timing guarantees in single-GPU real-time embedded systems. Using... Read more
Key finding: This study applies GPU acceleration to deep learning models for drug-target interaction prediction tasks pertinent to COVID-19. The DNN implementation leverages GPUs’ embarrassingly parallel computations during feed-forward... Read more
Key finding: This paper proposes a novel unsupervised weed detection method leveraging multispectral UAV imagery and the PC/BC-DIM neural network to generate fused saliency maps without requiring large labeled datasets or computationally... Read more
4. What are the architectural design considerations and comparative analyses for GPU programming models on modern supercomputers?
With the increasing reliance on GPUs in high-performance computing (HPC) facilities such as pre-exascale and exascale systems, this research area examines the performance tradeoffs, ease of use, portability, and optimization techniques of various GPU programming models. It involves evaluating vendor-supported languages (e.g., CUDA, HIP), directive-based models (OpenMP, OpenACC), and other abstractions (SYCL, Kokkos), especially focusing on AMD and NVIDIA GPUs in production supercomputers. Insights gained inform best practices for efficient GPU utilization and software portability across diverse HPC architectures.
Key finding: This work performs a comprehensive evaluation of widely used GPU programming models—CUDA, HIP, OpenMP offloading, OpenACC, hipSYCL, Kokkos, and Alpaka—on the LUMI supercomputer’s AMD MI250X GPUs, compared against NVIDIA... Read more
5. How can GPU memory access and data transfer mechanisms be enhanced using neural networks and advanced DMA controllers to improve multimedia and computing system performance?
GPU performance is often bottlenecked by inefficient memory access and data transfer between host and device. This research direction develops intelligent direct memory access (DMA) controllers leveraging back-propagation neural networks and advanced adaptive data placement to optimize memory channel usage, support high bandwidth transfers, and reduce power consumption. Applications include multimedia processing and heterogeneous computing involving GPU-FPGA systems, showing performance gains and reduced latency crucial for high-throughput GPU workloads.
Key finding: This work proposes a back-propagation algorithm (BPA) based neural network to dynamically control the DMA engine for multimedia applications on GPUs. By training the network to optimize gradient loss and adapting to various... Read more