Graduate by Hilton Annapolis126 West Street, Annapolis, MD, 21401
7:00 – 8:30 pm
7:00 – 8:00 am
8:00 – 8:15 am
Dave Mountain
8:15 – 8:30 am
Candy Culhane
8:30 – 9:15 am
Keynote
With the end of Moore’s Law and Dennard scaling, computing’s “free lunch” is over. Architectural innovation is now our primary lever for massive performance gains, yet we are tackling it with an outdated approach. Today’s human-driven teams explore less than 0.001% of a practically infinite design space. Relying on human intuition alone has become a severe bottleneck in an era where architectures must deliver 2x performance on the exact same transistor budget. This keynote explores the transition to automated idea factories. At the core of this shift is advanced data analytics, which transforms massive streams of system telemetry into the actionable intelligence that fuels our models. Powered by continuous machine learning and multi-tiered evaluation pipelines, these automated systems generate and evaluate thousands of candidate architectures weekly — compressing traditional design cycles from double-digit months down to single-digit weeks.
Karu Sankaralingram, NVIDIA
Session 1: Problems and Challenges
Dave Mountain, Moderator
9:15 – 10:00 am
Increasingly, people with no view inside the black box are stringing together AI and analytics systems in high-consequence settings. Who sets the rules and who answers when they fail? Washington, Beijing, and Rome are offering three different answers: a White House pressing Congress to override state AI laws while betting on AI-driven science; a Chinese Communist Party racing for compute and standards; and a Pope — a former math major — whose AI encyclical puts human dignity and responsibility at the center. Devan Patel maps the state of play and what it asks of the people building these systems.
Devan Patel, Crux Advisory Group and Rainey Center
10:00 – 10:15 am
10:15 – 11:00 am
What are the actual capabilities of AI in research, especially for foundational research, stripped of hype? This talk will cover independent evaluations on math research problems that show there are still limits, despite some sensational successes. It will cover how AI actually works, considering a functional view of what LLMs are, and discuss the mounting stresses on scholarly publishing. The impact of AI on research could go many ways, and we as a field have some control over this. Part of the answer might me more independent assessments, so that the view of the AI companies doesn’t dominate. The wish is for foundational research to be grander than today, valuing fewer, higher-quality contributions (versus easily generated quantity), and motivated by intense curiosity and passion. But this is just the beginning of the discussion!
Tammy Kolda, MathSci.AI
11:00 – 11:45 am
Financial intelligence analysis, illicit finance investigations, fraud prevention, and other related disciplines have grown in complexity as criminals, money launderers, and illicit finance actors have adopted increasingly sophisticated methods to evade detection by traditional enforcement means and transfer value through financial systems. The existing financial system’s anti-money laundering and fraud regimes are built on a system that traditionally detected illicit financial flows from the perspective of the identities of transacting parties. Anonymity-enhancing technologies and increasingly sophisticated and complex financial systems have moved the state of illicit finance beyond identity-centric detection; approaches that keep pace with innovation need to transition to activity-based anomaly detection methods. Observation of anomalous patterns of behaviors within transaction datasets and across disconnected data domains needs to become the new foundational approach. This transformation presents challenges in the scale of analytics, data storage and architecture, and computing power. Artificial intelligence models that rely on multiple, disparate data sources of varying levels of data quality need solid training and engineering to provide valuable insights to policymakers and regulators.
Mike Morrison, Booz Allen Hamilton
11:45 am – 12:30 pm
Panel Discussion
12:30 – 1:45 pm
Session 2: Systems and Applications I
Brad Spiers, Moderator
1:45 – 2:30 pm
CPUs and GPUs keep automata in memory, enabling flexible reconfiguration but relying on data-dependent memory accesses. Spatial accelerators provide more regular execution by mapping automata into hardware, but tie capacity and reconfiguration to physical resources. We present a middle ground based on learned range matching, where transitions are encoded as multidimensional ranges over state and input space. A compact learned model predicts the location of the matching range, followed by an exact correction step: learning affects performance, never correctness. We implement this model as a deeply pipelined FPGA architecture that interleaves hundreds of streams while keeping the automaton entirely in memory and reconfigurable without FPGA synthesis. Applied to inline policy processing for AI textual data, the system sustains up to 1.5 GB/s of application payload with sub-10-ms P99 latency, while inspecting every byte against thousands of patterns.
Alon Rashelbach, NOL8
2:30 – 3:15 pm
Modern technical sensor stream processing and analysis are fundamentally constrained by the limitations of deterministic logic when confronted with uncertainty, obfuscation, and massive data volume. To address this grand challenge, this presentation details the development of FIDO, a cognitive neuro-symbolic graph processing framework, and its hardware execution layer, a printed photonic fractal inference engine. Moving beyond binary paradigms, the FIDO chip utilizes light-based fractal geometries to natively execute probabilistic graph analysis, seamlessly merging neural network flexibility with symbolic reasoning. We will present the hardware-software co-design that allows FIDO to map complex, non-deterministic relationships — endemic to massive intercept datasets — directly into optical pathways. By leveraging printed photonics, this architecture not only achieves extreme scalability, but also drastically reduces Size, Weight, and Power (SWaP) requirements. The discussion will focus on the implications of this neuro-symbolic approach for the CLASSIC community, specifically how transitioning from deterministic silicon to probabilistic photonic fractals can unlock real-time, scalable anomaly detection and complex network analysis in heavily constrained deployment environments.
Greg Compton, AbraWorks
3:15 – 3:30 pm
Break
3:30 – 4:15 pm
This talk will summarize the Flow-Optimized Reconfigurable Zones of Acceleration (FORZA) project, undertaken as part of the IARPA AGILE program. FORZA enables system-wide asynchronous movement of computation to data through a Fine-grained Asynchronous Bulk-Synchronous Parallelism (FA-BSP) model. Using software-hardware co-design, the project combines an actor-based programming system with a novel near-memory architecture, while also supporting commercial off-the-shelf (COTS) HPC systems and modern runtimes for Agentic AI. A full FORZA system is envisioned to support more than 30 million hardware threads across 20 racks. Projected performance gains on AGILE benchmarks span multiple orders of magnitude, including speedups of over 350,000× for graph construction and 180,000× for ego-graph construction and GCN training. Implementations of FA-BSP on COTS systems have also achieved substantial gains, earning SCALE Challenge Awards in 2023 and 2024. These results illustrate FORZA’s potential to deliver transformative performance improvements in large-scale data analytics on the FORZA architecture as well as significant speedups on COTS HPC platforms.
Vivek Sarkar, Georgia Institute of Technology
4:15 – 5:00 pm
Panel Discussion
5:00 – 6:00 pm
Break
6:00 – 7:30 pm
Dinner Speaker
Breakthroughs happen when we find a conduit to channel potential into action. But how do we channel the massive rivers of data streaming down from Earth-orbiting satellites into actionable, fit-for-purpose information? As well, real-world decisions require complex combinations of data, models, and even political considerations. How do we support those decision-makers, especially ones who lack technical knowledge of the capabilities? Can AI help us? In this talk we explore the promise and challenge of using Earth-information to respond to disasters, improve agriculture, and manage the screwworm invasion.
Tom Wagner, NASA
7:30 – 9:00 pm
Reception
7:00 – 8:30 am
Breakfast
8:30 – 9:15 am
Keynote
Automated scientific discovery requires a closed-loop system that iterates through hypothesis generation, computation, experiment, and analysis, improving with each cycle. This talk presents a layered set of building blocks for constructing such loops, drawn from the Genesis Mission and AI efforts at Sandia. We will cover the Genesis Mission’s Science and Security Platform, which federates AI infrastructure across DOE national laboratories. We will then describe the services, libraries, and workflows needed to develop a discovery loop that can sustain an autonomous scientific discovery campaign. The core of such a campaign depends on foundation models for science and engineering; we will highlight generative models for these use cases. Agentic workflows can be built on top of the infrastructure and models. We will present a recent workflow built by our team for designing, manufacturing, and testing a flight test vehicle. Finally, the talk will close with open research problems: evaluating the reasoning capabilities of scientific models and equation discovery. The goal is to develop a common set of reusable components for scientific discovery that lets teams assemble discovery loops for new domains without rebuilding the stack.
Siva Rajamanickam, Sandia National Laboratories
Session 3: Systems and Applications II
David Haglin, Moderator
9:15 – 10:00 am
Because sparse graph queries involve irregular, fine-grained data dependencies, performance at large scale is dominated by data access latency rather than by computation rate or bandwidth. Although latency hiding through asynchrony and message aggregation has proved useful, these approaches remain constrained by networks typically optimized for bulk data transfers. This talk examines how special-purpose hardware can address that latency constraint across a range of graph queries. The approach combines reusable FPGA kernels with a network designed for fine-grained, low-latency communication. Relational algebra guides query optimizations that exploit these kernels while accounting for communication across cards, allowing hardware specialization to support classes of queries rather than individual queries. Using selected graph queries, we draw on analytic modeling, simulation, and experimental results from FPGAs connected by Data Vortex to examine the potential of this approach relative to distributed-memory CPU systems.
Andrew Lumsdaine, Pacific Northwest National Laboratory
10:00 – 10:15 am
Break
10:15 – 11:00 am
This talk revisits a classical problem that has received decades of attention: sparse matrix-dense vector multiplication (SpMV) on irregular graphs such as those in social network analysis and other subfields of data science. We will review the properties of such graphs, as well as some existing HPC solutions. We will introduce a new algorithm for SpMV on unstructured graphs that exploits dataflow hardware such as the Cerebras wafer-scale engine. We emphasize the non-procedural thinking involved in the design of this algorithm, and we obtain runtime performance on Cerebras CS3 that is superior to that of cuSparse on modern GPUs. This talk covers joint work with Kasimir Gabert (Sandia National Labs), Ilya Sharapov (Cerebras), Tomas Oppelstrup (Cerebras), and Maya Taylor (Sandia National Labs).
Jon Berry, Sandia National Laboratories
11:00 – 11:45 am
The two mainstream architectures (GPUs: dense, regular, AI, and CPUs: catchall), fail to meet acceleration needs for high speed ingestion/transformation, irregular graph/sparse computation, and large-scale context for agentic AI. UpDown’s novel architecture and software deliver breakthrough performance on graph computing: 100x (CPU) and 30x (GPU). Efficient exploitation of fine-grained parallelism (1-cycle thread creation, 10 inst invocations, and massive HBM bandwidth), and high-level map-reduce programming enable strong scaling on small AND large problems, achieving 10,000-fold overall speedups on BFS, Pagerank, CC, K-truss, Louvain, K-cores, and more. Large UpDowns exceed supercomputer performance at 10x lower power. Small UpDown systems can deliver TeraTEPS of performance. UpDown enables complex GraphRAG with sophisticated dynamic graphs for agentic AI.
Streaming analytics and complex workflows also require high speed ingestion for flexible composition across data systems and heterogeneous data. UpDown delivers breakthrough ingestion performance of >5 billion records/second/node scaling to >50 Tera-records/second. This is a thousand times faster than existing commercial systems, making flexible, complex composition a practical capability. UpDown is the product of four years of IARPA AGILE-funded research, and we are building ASIC designs in a spinout from the University of Chicago. chupdown.com UpDown’s K-truss implementation won an MIT Graph Challenge Innovation Award in 2025.
Andrew Chien, The University of Chicago and Chicago UpDown Computing
11:45 am – 12:30 pm
Panel Discussion
12:30 – 5:00 pm
Free Time: Lunch and early dinner (on your own)
2:00 pm
Naval Academy (Walking) Tour
5:00 – 6:30 pm
Random Access
Katie Wolf
6:30 – 8:30 pm
Poster Reception
Compiler-Driven Message Aggregation for One-Sided Communications Using OpenSHMEM
Joseph Cowles, Texas Tech University
Mitigating Incast Congestion with User-Space, Node-Based Flow Control
Noel Gomillion, Texas A&M University
Secure, Performant Tensor Computations Using Rust from Python
Kimmie Harding, Sandia National Laboratories
Implications of Reversible Control Flow in Quantum Optimization
Cameron Ibrahim, University of Delaware
UpDown: Efficient Manycore Based on Many Threading and Scalable Memory Parallelism
Andronicus Rajasukumar, University of Chicago
7:00 – 8:30 am
8:30 – 9:15 am
Keynote
AI scientist agents promise to revolutionize how science is conducted, substantially increasing scientists’ productivity and accelerating discovery and innovation. By connecting scientific data, domain-specific AI models, simulations, and experiments, these agents can orchestrate research workflows — from hypothesis exploration to analysis — while enabling researchers to concentrate on scientific insight and creative inquiry. This keynote presents RIKEN’s vision for AI for Science and Industry, illustrated through Fugaku-enabled advances in cardiac digital twins, weather and climate modelling, scientific imaging, and autonomous drug discovery. We will outline the infrastructure supporting this transformation, from RIKYU and quantum-HPC hybrid platforms to FugakuNEXT, where hardware, software, and applications are being co-designed for greater performance and energy efficiency. Japan–U.S. collaboration through the Genesis Mission further extends this vision toward a shared scientific ecosystem that accelerates research cycles and turns discoveries into technological and industrial innovation.
Satoshi Matsuoka, RIKEN
Session 4: Institutions, Systems, and the Value of Research
Laura Monroe, Moderator
9:15 – 10:00 am
Agentic AI promises to turn scientific workflows from pipelines humans orchestrate into systems that plan, delegate, call tools, and act. At a national laboratory, that autonomy must operate within shared supercomputers, batch schedulers, finite allocations, and failure-prone long-running jobs, where every action must remain observable and auditable. This talk will describe Argonne’s experience moving agentic biomolecular design from prototype to production on Polaris and Aurora.
This talk will focus on StructBioReasoner, a multi-agent framework for biomolecular discovery and design. Agents generate and refine peptide binders against targets that resist conventional structure-based methods, including intrinsically disordered proteins and glycosylated viral surface proteins. For enzyme engineering, a discovery agent reasons from a desired transformation to candidate enzymes, prioritizing families and screening for substrate compatibility. Underneath is an execution stack (MCP, Globus Compute, Parsl, Academy) that lets long-running agents submit, monitor, and recover HPC workloads without persistent interactive sessions. We will discuss what makes this governable: bounded autonomy, safety screening of agent-generated code, resource accounting, fault recovery, and human oversight.
Matt Sinclair, Argonne National Laboratory
10:00 – 10:15 am
Break
10:15 – 11:00 am
Money laundering has huge worldwide impact. The UN estimates that every year 2–5% of global GDP is laundered, about $800–$2,000 billion, which is used to support a broad variety of criminal enterprises and terrorist activity. Alas, the UN and others also estimate that only a small fraction of laundering is ever detected, e.g. 1%. This talk will outline why Anti Money Laundering (AML) currently fares so poorly. The talk will then outline how the emergence of new graph algorithms, AI, highly capable hardware, and high-quality synthetic data offer the opportunity for a dramatic rethinking of how AML is practiced.
Erik Altman, IBM
11:00 – 11:30 am
Panel Discussion
11:30 – 12:30 pm
There is a need for better ways to measure and show the value of HPC and AI in science and engineering. The cost of systems is growing quickly and often requires major justification from funding bodies. This panel will start with a description of a new project to measure the value of science and engineering and ways to show the ROI from investments in HPC and AI for science. Sites, nations, and companies around the world are expanding their investments in HPC/AI at a rate not seen before, and will require “proof” of the value of these investments at some point in the future. The panel will then discuss different ways to measure and show the value of science.
Earl Joseph, Hyperion Research
Candy Culhane, Los Alamos National Laboratory
Giri Chukkapalli, NVIDIA
12:30 – 1:45 pm
Lunch
Session 5: What’s Next
Justin Newcomer, Moderator
1:45 – 2:30 pm
The end of Moore’s Law with nano-scale feature size and bounded clock rates requires revolutionary approaches to future computer architecture in order to expand performance growth. In spite of the potential promise of Quantum Computing and neuromorphic processing, these uncertain concepts are many years in the future, if ever. The Active Memory Architecture is a non-von Neumann memory-centric structure that simplifies organization, expands diversity of parallelism, dramatically increases concurrency of communication, and supports asynchrony of operation to adjust for variation of latencies. Together, these attributes reinforce algorithms such as graph processing.
Thomas Sterling, University of Texas at Austin and TACC
2:30 – 3:15 pm
The healthcare industry accounts for a very large part of the GDP in the United States and yet, it lags behind many other industries in terms of utilization of computational resources to improve quality. In this talk we will discuss multiple aspects of the tremendous opportunity to revolutionize healthcare delivery and contrast it with the very unique and specific issues that make this a daunting challenge. Specifically we will discuss, with real life examples, opportunities and needs in efficient delivery of healthcare, personalized medicine, drug discovery, and population level modeling. Throughout, we will highlight examples from the current state of these advancements in the medical world and ask the following question: what is keeping us from moving forward? We will explore important concepts such as the challenges associated with modeling biological systems with limited quantitative data, legal frameworks, the structure/business model of the industry, public/ethical boundaries, and the irrational nature of medical decision making. The overall objective is, in this short time, to communicate a flavor of the specific challenges and opportunities that have to be kept in mind when translating computational breakthroughs into the U.S medical industry.
Farbod “Nick” Rahaghi, University of California, Los Angeles
3:15 – 3:30 pm
Break
3:30 – 4:15 pm
We are developing an agentic digital twin framework for Earth system tipping point discovery that uses AI agents to orchestrate scientific investigation across heterogeneous physics-based and AI models. The framework coordinates reduced-order dynamical models, generative AI, forecasting models, and analysis tools to explore large parameter spaces, identify previously unexplored dynamical regimes, characterize critical transitions, and guide subsequent experiments. Rather than executing a fixed computational workflow, specialized agents reason over intermediate results to determine what analyses, models, and experiments should be performed next, enabling a closed-loop approach to scientific discovery. This work explores how agentic AI can move digital twins beyond simulation and prediction toward adaptive scientific systems that integrate modeling, experimentation, and reasoning to investigate complex dynamical systems at scale.
Jennifer Sleeman, Johns Hopkins University
4:15 – 5:00 pm
Panel Discussion
5:00 – 5:15 pm
Closing Remarks and Adjourn
Richard Murphy
5:15 – 6:30 pm
Closing Reception