The surge in scientific literature obscures breakthroughs and hinders the discovery of new research paths. We propose an artificial intelligence (AI) powered framework using large language models (LLMs) and knowledge graphs (KGs) to automate parts of scientific discovery, focusing on energy-efficient AI circuits. Our hybrid approach combines LLMs, structured data, and ontology-based reasoning to construct a comprehensive knowledge graph that integrates insights across computational neuroscience, spiking neuron models, learning rules, architectural motifs, and neuromorphic device technologies. This multi-domain representation enables the generation of hypotheses that connect biological function with implementable, energy-efficient hardware architectures. Using KG embeddings and graph neural networks, the framework generates hypotheses for novel circuits, validates them through optimization on exascale HPC systems, and with tools like SuperNeuro and Fugu, the most promising designs will be prototyped in hardware. This open-source system aims to accelerate discoveries and bridging neuroscience with hardware innovation, drive collaboration, and unlock new opportunities in low-power AI computing.

Multi-spike models of synaptic plasticity, such as the triplet and suppression spike-timing-dependent plasticity (STDP) rules, exhibit better alignment with neurophysiological data in the brain compared to the pair-based STDP rule. Previous studies have empirically shown that the pair-based STDP rule can detect spatiotemporal spike patterns hidden in equally dense distractor spike trains in an unsupervised manner. However, it fails to detect spike patterns influenced by jitter noise. Given that spiking neural networks (SNNs) exhibit variability in generated spike trains in response to the same inputs, it becomes imperative to have learning rules capable of detecting spike patterns even in the presence of jitter noise. In this study, we introduce a simplified suppression-based STDP rule that demonstrates significantly enhanced tolerance to jitter in spike patterns compared to the pair-based STDP rule. Unlike the ideal suppression STDP rule, characterized by an exponential learning window and requiring high-resolution synapses, the simplified rule limits the synaptic efficacy update to a single bit at any given instant. Moreover, it employs 4-bit fixed-point synapses, facilitating straightforward implementation in neuromorphic hardware.

In many neuromorphic workflows, simulators play a vital role for important tasks such as training spiking neural networks, running neuroscience simulations, and designing, implementing, and testing neuromorphic algorithms. Currently available simulators cater to either neuroscience workflows (e.g., NEST and Brian2) or deep learning workflows (e.g., BindsNET). Problematically, the neuroscience-based simulators are slow and not very scalable, and the deep learning-based simulators do not support certain functionalities that are typical of neuromorphic workloads (e.g., synaptic delay). In this paper, we address this gap in the literature and present SuperNeuro, which is a fast and scalable simulator for neuromorphic computing capable of both homogeneous and heterogeneous simulations as well as GPU acceleration. We also present preliminary results that compare SuperNeuro to widely used neuromorphic simulators such as NEST, Brian2, and BindsNET in terms of computation times. We demonstrate that SuperNeuro can be approximately 10×--300× faster than some of the other simulators for small sparse networks. On large sparse and large dense networks, SuperNeuro can be approximately 2.2×--3.4× faster than the other simulators, respectively.

The new reality of exascale computing faces many challenges in achieving optimal performance on large numbers of nodes. A key challenge is the efficient utilization of the message-passing interface (MPI), a critical component for process communication. This paper explores communication optimization strategies to harness the GPU-accelerated architectures of these supercomputers. We focus on MPI applications where processors form a two-dimensional process grid, a common arrangement in applications involving dense matrix operations. This configuration offers a unique opportunity to implement innovative strategies to improve performance and maintain effective load distribution. We study two applications— Dist-FW (Apsp:all-pair-shortest-path) and HPL-MxP (LU factorization with Mixed precision)—on two accelerated systems: Summit (IBM Power and NVIDIA V100) and Frontier (AMD EPYC and MI250X). These supercomputers are operated by the Oak Ridge Leadership Computing Facility (OLCF) and are currently ranked #1 and #5 on the Top500 list. We show how to scale up both applications to exascale levels and tackle the MPI challenges related to implementation, synchronization, and performance. We also compare the performance of several communication strategies at an unprecedented scale. Accurately predicting application performance becomes crucial for cost reduction as the computational scale grows. To address this, we suggest a hyperbolic model as a better alternative to the traditional one-sided asymptotic model for predicting future application performance at such large scales.

In the upcoming decade, deep learning may revolutionize the natural sciences, enhancing our capacity to model and predict natural occurrences. This could herald a new era of scientific exploration, bringing significant advancements across sectors from drug development to renewable energy. To answer this call, we present DeepSpeed4Science initiative (deepspeed4science.ai) which aims to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. By leveraging DeepSpeed's current technology pillars (training, inference and compression) as base technology enablers, DeepSpeed4Science will create a new set of AI system technologies tailored for accelerating scientific discoveries by addressing their unique complexity beyond the common technical approaches used for accelerating generic large language models (LLMs). In this paper, we showcase the early progress we made with DeepSpeed4Science in addressing two of the critical system challenges in structural biology research.

Neuromorphic computers emulate the human brain while being extremely power efficient for computing tasks. In fact, they are poised to be critical for energy-efficient computing in the future. Neuromorphic computers are primarily used in spiking neural network–based machine learning applications. However, they are known to be Turing-complete, and in theory can perform all general-purpose computation. One of the biggest bottlenecks in realizing general-purpose computations on neuromorphic computers today is the inability to efficiently encode data on the neuromorphic computers. To fully realize the potential of neuromorphic computers for energy-efficient general-purpose computing, efficient mechanisms must be devised for encoding numbers. Current encoding mechanisms (e.g., binning, rate-based encoding, and time-based encoding) have limited applicability and are not suited for general-purpose computation. In this paper, we present the virtual neuron abstraction as a mechanism for encoding and adding integers and rational numbers by using spiking neural network primitives. We evaluate the performance of the virtual neuron on physical and simulated neuromorphic hardware. We estimate that the virtual neuron could perform an addition operation using just 23 nJ of energy on average with a mixed-signal, memristor-based neuromorphic processor. We also demonstrate the utility of the virtual neuron by using it in some of the μ -recursive functions, which are the building blocks of general-purpose computation.

Neuromorphic computing technology continues to make strides in the development of new algorithms, devices, and materials. In addition, applications have begun to emerge where neuromorphic computing shows promising results. However, numerous barriers to further development and application remain. In this work, we identify several science areas where neuromorphic computing can either make an immediate impact (within 1 to 3 years) or the societal impact would be extremely high if the technological barriers can be addressed. We identify both opportunities and hurdles to the development of neuromorphic computing technology for these areas. Finally, we discuss future directions that need to be addressed to expand both the development and application of neuromorphic computing.

The 2021 Smoky Mountains Computational Sciences and Engineering Conference enlists scientists from across Oak Ridge National Laboratory (ORNL) and industry to be data sponsors and help create data analytics and edge computing challenges for eminent datasets in a variety of scientific domains. This work describes the significance of each of the eight datasets and their associated challenge questions. The challenge questions for each dataset were required to cover multiple difficulty levels. An international call for participation was sent to students, asking them to form teams of up to six people and apply novel data analytics and edge computing methods to solve these challenges.

We are motivated by newly proposed methods for mining large-scale corpora of scholarly publications (e.g., full biomedical literature), which consists of tens of millions of papers spanning decades of research. In this setting, analysts seek to discover relationships among concepts. They construct graph representations from annotated text databases and then formulate the relationship-mining problem as an all-pairs shortest paths (APSP) and validate connective paths against curated biomedical knowledge graphs (e.g., Spoke). In this context, we present Coast (Exascale Communication-Optimized All-Pairs Shortest Path) and demonstrate 1.004 EF/s on 9,200 Frontier nodes (73,600 GCDs). We develop hyperbolic performance models (HYPERMOD), which guide optimizations and parametric tuning. The proposed Coast algorithm achieved the memory constant parallel efficiency of 99% in the single-precision tropical semiring. Looking forward, Coast will enable the integration of scholarly corpora like PubMed into the Spoke biomedical knowledge graph.

Neuromorphic computing is a non-von Neumann computing paradigm that performs computation by emulating the human brain. Neuromorphic systems are extremely energy-efficient and known to consume thousands of times less power than CPUs and GPUs. They have the potential to drive critical use cases such as autonomous vehicles, edge computing and internet of things in the future. For this reason, they are sought to be an indispensable part of the future computing landscape. Neuromorphic systems are mainly used for spike-based machine learning applications, although there are some non-machine learning applications in graph theory, differential equations, and spike-based simulations. These applications suggest that neuromorphic computing might be capable of general-purpose computing. However, general-purpose computability of neuromorphic computing has not been established yet. In this work, we prove that neuromorphic computing is Turing-complete and therefore capable of general-purpose computing. Specifically, we present a model of neuromorphic computing, with just two neuron parameters (threshold and leak), and two synaptic parameters (weight and delay). We devise neuromorphic circuits for computing all the μ-recursive functions (i.e., constant, successor and projection functions) and all the μ-recursive operators (i.e., composition, primitive recursion and minimization operators). Given that the μ-recursive functions and operators are precisely the ones that can be computed using a Turing machine, this work establishes the Turing-completeness of neuromorphic computing.

Neuromorphic computing offers the opportunity to implement extremely low power artificial intelligence at the edge. Control applications, such as autonomous vehicles and robotics, are also of great interest for neuromorphic systems at the edge. It is not clear, however, what the best neuromorphic training approaches are for control applications at the edge. In this work, we implement and compare the performance of evolutionary optimization and imitation learning approaches on an autonomous race car control task using an edge neuromorphic implementation. We show that the evolutionary approaches tend to achieve better performing smaller network sizes that are well-suited to edge deployment, but they also take significantly longer to train. We also describe a workflow to allow for future algorithmic comparisons for neuromorphic hardware on control applications at the edge.

Neuromorphic computers perform computations by emulating the human brain and are expected to be indispensable for energy-efficient computing in the future. They are primarily used in spiking neural network-based machine learning applications. However, neuromorphic computers are unable to preprocess data for these applications. Currently, data is preprocessed on a CPU or a GPU-this incurs a significant cost of transferring data from the CPU/GPU to the neuromorphic processor and vice versa. This cost can be avoided if preprocessing is done on the neuromorphic processor. To efficiently preprocess data on a neuromorphic processor, we first need an efficient mechanism for encoding data that can lend itself to all general-purpose preprocessing operations. Current encoding approaches have limited applicability and may not be suitable for all preprocessing operations. In this paper, we present the virtual neuron as a mechanism for encoding integers and rational numbers on neuromorphic processors. We evaluate the performance of the virtual neuron on physical and simulated neuromorphic hardware and show that it can perform an addition operation using 23 nJ of energy on average using a mixed-signal, memristor-based neuromorphic processor. The virtual neuron encoding approach is the first step in preprocessing data on a neuromorphic processor.

Graph convolutional networks have risen in popularity in recent years to tackle problems that are naturally represented as graphs. However, real-world graphs are often sparse, which means that implementing them on traditional accelerators such as graphics processing units (GPUs) can lead to inefficient utilization of the hardware. Spiking neuromorphic computers natively implement network-like computation and have been shown to be successful at implementing certain types of graph computations. In this work, we evaluate the use of a simulated network of spiking neurons to perform semi-supervised learning on graph data using only the graph structure. We demonstrate that our neuromorphic approach provides comparable results to graph convolutional network results, andwe discuss the opportunities for using neuromorphic computers for this task in the future.

Neuromorphic computing systems offer many attractive qualities for real-world applications, including extremely low power and relatively fast computation. There are many new and emerging devices and materials that are being evaluated by the research community for neuromorphic deployment. However, as this is still an active research field, there are relatively few application-ready, full-scale neuromorphic systems that are available to the community. As such, much of neuromorphic research is currently being performed via a bottom-up approach, in which the hardware is defined first and then applications are mapped onto the hardware after. The lack of complete neuromorphic solutions, including algorithms and software, as well as full-scale neuromorphic implementation, does not allow for application-deployment and evaluation of neuromorphic computing solutions for real-world applications. To address this issue, we have developed a complete neuromorphic framework (Figure 1), including algorithms for designing and training spiking neural networks (SNNs) for neuromorphic deployment, a common software interface to a variety of neuromorphic backends, and a field-programmable gate array (FPGA)-based neuromorphic implementation to study architectural requirements of

Neuromorphic computers offer the opportunity for low-power, efficient computation. Though they have been primarily applied to neural network tasks, there is also the opportunity to leverage the inherent characteristics of neuromorphic computers (low power, massive parallelism, collocated processing and memory) to perform non-neural network tasks. Here, we demonstrate how an approach for performing sparse binary matrix-vector multiplication on neuromorphic computers. We describe the approach, which relies on the connection between binary matrix-vector multiplication and breadth first search, and we introduce the algorithm for performing this calculation in a neuromorphic way. We validate the approach in simulation. Finally, we provide a discussion of the runtime of this algorithm and discuss where neuromorphic computers in the future may have a computational advantage when performing this computation.

We present an optimized Floyd-Warshall (Floyd-Warshall) algorithm that computes the All-pairs shortest path (APSP) for GPU accelerated clusters. The Floyd-Warshall algorithm due to its structural similarities to matrix-multiplication is well suited for highly parallel GPU architectures. To achieve high parallel efficiency, we address two key algorithmic challenges: reducing high communication overhead and addressing limited GPU memory. To reduce high communication costs, we redesign the parallel (a) to expose more parallelism, (b) aggressively overlap communication and computation with pipelined and asynchronous scheduling of operations, and (c) tailored MPI-collective. To cope with limited GPU memory, we employ an offload model, where the data resides on the host and is transferred to GPU on-demand. The proposed optimizations are supported with detailed performance models for tuning. Our optimized parallel Floyd-Warshall implementation is up to 5x faster than a strong baseline and achieves 8.1 PetaFLOPS/sec on 256~nodes of the Summit supercomputer at Oak Ridge National Laboratory. This performance represents 70% of the theoretical peak and 80% parallel efficiency. The offload algorithm can handle 2.5x larger graphs with a 20% increase in overall running time.

A major challenge in machine learning is the computational expense of training these models. Model training can be viewed as a form of optimization used to fit a machine learning model to a set of data, which can take up significant amount of time on classical computers. Adiabatic quantum computers have been shown to excel at solving optimization problems, and therefore, we believe, present a promising alternative to improve machine learning training times. In this paper, we present an adiabatic quantum computing approach for training a linear regression model. In order to do this, we formulate the regression problem as a quadratic unconstrained binary optimization (QUBO) problem. We analyze our quantum approach theoretically, test it on the D-Wave 2000Q adiabatic quantum computer and compare its performance to a classical approach that uses the Scikit-learn library in Python. Our analysis shows that the quantum approach attains up to 2.8x speedup over the classical approach on larger datasets, and performs at par with the classical approach on the regression error metric.

Neuromorphic computing is poised to become a promising computing paradigm in the post Moore’s law era due to its extremely low power usage and inherent parallelism. Spiking neural networks are the traditional use case for neuromorphic systems, and have proven to be highly effective at machine learning tasks such as control problems. More recently, neuromorphic systems have been applied outside of the arena of machine learning, primarily in the field of graph algorithms. Neuromorphic systems have been shown to perform graph algorithms faster and with lower power consumption than their traditional (GPU/CPU) counterparts, and are hence an attractive option for a co-processing unit in future high performance computing systems, where graph algorithms play a critical role. In this paper, we present a neuromorphic implementation of cycle detection, odd cycle detection, and the Ford-Fulkerson max-flow algorithm. We further evaluate the performance of these implementations using the NEST neuromorphic simulator by using spike counts and simulation time as proxies for energy consumption and run time. In addition to gains inherent in neuromorphic systems, we show that within the neuromorphic implementations early stopping criteria can be implemented to further improve performance.