Disentanglement of Latent Representations through Adversarial Covariance Minimization CSMD ORNL Computer Science and Mathematics Divsion
Disentanglement of Latent Representations through Adversarial Covariance Minimization
Self-Supervised Novelty Detection for Continual Learning CSMD Computer Science and Mathematics Division ORNL
Self-Supervised Novelty Detection for Continual Learning

Architectures and Performance group

At Oak Ridge National Laboratory, the Computing and Computational Sciences Directorate’s (CCSD’s) Architectures and Performance group is transforming how scientists harness the power of supercomputers and advanced computing systems by leveraging cutting-edge neuromorphic, analog, quantum, and heterogeneous computing paradigms. This work addresses a crucial challenge in modern scientific research: making increasingly complex computing systems more accessible and efficient for researchers across different fields.

The group’s mission is to create a collaborative microelectronics center of excellence dedicated to energy-efficient and performance-forward co-design. Leveraging its strategic position, the group advances research across novel computing paradigms, runtime systems, high-level programming models, and high-performance computing (HPC) and machine-learning applications to optimize both performance and energy use. The group’s research in emerging computing paradigms helps advance the US Department of Energy’s scientific objectives.

The group’s flagship innovation is the Intelligent Runtime System (IRIS), which is a sophisticated heterogeneous runtime framework that enables scientific applications to run seamlessly across various computing systems—from smartphones to supercomputers. As a recipient of an R&D 100 award in 2024, the IRIS team has enhanced the framework with cutting-edge machine learning capabilities, a bespoke heterogenous memory handler, automatic data orchestration, automatic dataflow optimizations, and a mechanism to use graph neural networks to intelligently optimize how computational tasks are scheduled and executed. These advancements help scientists focus on their research by providing them with a runtime that supports easier programmability, higher productivity, and improved performance without requiring low-level or accelerator-specific coding expertise.

The group has provided outstanding architectures for ultralow-latency digital-twin machine-learning solutions on field-programmable gate arrays, including solutions that are now deployed at the Spallation Neutron Source and the Center for Nanophase Materials Sciences user facilities. The group has also advanced neuromorphic research through the Neuro-Spark architecture with an ultralow latency (~50 nanoseconds) prediction solution for the high-PT and low-PT classification of charge particles for the Large Hadron Collider. The group is now working on an ASIC tapeout of this solution.

Additionally, the group is actively conducting co-design research on emerging memory technologies, including ReRAM (resistive RAM) and ECRAM (electrochemical RAM), and investigating advanced processing-in-memory strategies for HPC and machine learning applications.

The group’s practical impact is demonstrated through several key projects:

  • Sub-microsecond spiking neural network solution (Neuro-Spark architectures)
  • ReSpike co-design framework for evaluating spiking neural networks on ReRAM
  • MatRIS multilevel heterogeneous math library for performance portability on heterogeneous computing architectures
  • Abisko co-designed architecture for spiking neural networks using novel neuromorphic materials
  • Improved performance of the Spallation Neutron Source facility through errant beam detection by using machine learning–based optimization to directly contribute to more efficient clean energy research

What makes this group’s approach particularly valuable is the combination of practical solutions with forward-thinking innovation. Rather than creating isolated tools, the team is building an integrated ecosystem that bridges the gap between complex computing hardware and the scientists who use it. The group’s work spans from improving basic programming tools to developing sophisticated AI-driven systems that automatically optimize performance.

Additionally, the research group also offers access to the Experimental Computing Laboratory (ExCL), which is built for computer science research and supports CCSD by providing a stable and secure computing environment. At a time when heterogeneity defines the path forward, ExCL offers heterogeneous resources that scientists can use in their work, thereby enabling them to focus on their research instead of worrying about acquiring and maintaining the underlying hardware and infrastructure. The ExCL’s resources comprise a diverse set of processor, memory, and storage technologies. This hardware is continually adapted and upgraded to incorporate the latest architectures to ensure that ExCL users have access to the state of the art.

Looking ahead, the group is positioned to play a crucial role in democratizing access to advanced computing resources. As supercomputers become more powerful and complex, the group’s tools and systems will be essential to ensuring that these resources can be effectively used by researchers working on solutions to the world’s most pressing challenges—from climate modeling to drug discovery.

The group’s vision of creating more intelligent and accessible computing systems promises to accelerate scientific discovery across disciplines, ultimately contributing to breakthroughs that benefit humanity.