Energy efficiency of training and inferencing with large neural network models is a critical challenge facing the future of sustainable large-scale machine learning workloads. This paper introduces an alternative strategy, called phantom parallelism, to minimize the net energy consumption of traditional tensor (model) parallelism, the most energy-inefficient component of large neural network training. The approach is presented in the context of feed-forward network architectures as a preliminary, but comprehensive, proof-of-principle study of the proposed methodology. We derive new forward and backward propagation operators for phantom parallelism, implement them as custom autograd operations within an end-to-end phantom parallel training pipeline and compare its parallel performance and energy-efficiency against those of conventional tensor parallel training pipelines. Formal analyses that predict lower bandwidth and FLOP counts are presented with supporting empirical results on up to 256 GPUs that corroborate these gains. Experiments are shown to deliver$\sim 50 \%$reduction in the energy consumed to train FFNs using the proposed phantom parallel approach when compared with conventional tensor parallel methods. Additionally, the proposed approach is shown to train smaller phantom models to the same model loss on smaller GPU counts as larger tensor parallel models on larger GPU counts offering the possibility for even greater energy savings.

Emerging expert-specialized Mixture-of-Experts (MoE) architectures, such as DeepSeek-MoE, deliver strong model quality through fine-grained expert segmentation and large top-k routing. However, their scalability is limited by substantial activation memory overhead and costly all-to-all communication. Furthermore, current MoE training systems – primarily optimized for NVIDIA GPUs – perform suboptimally on non-NVIDIA platforms, leaving significant computational potential untapped. In this work, we present X-MoE, a novel MoE training system designed to deliver scalable training performance for next-generation MoE architectures. X-MoE achieves this via several novel techniques, including efficient padding-free MoE training with cross-platform kernels, redundancy-bypassing dispatch, and hybrid parallelism with sequence-sharded MoE blocks. Our evaluation on the Frontier supercomputer, powered by AMD MI250X GPUs, shows that X-MoE scales DeepSeek-style MoEs up to 545 billion parameters across 1024 GPUs – 10x larger than the largest trainable model with existing methods under the same hardware budget, while maintaining high training throughput.

Genome data from cancer patients represents relationships between the presence of a gene mutation and cancer occurrence in a patient. Different types of cancer in human are thought to be caused by combinations of two to nine gene mutations. Identifying these combinations through traditional exhaustive search requires the amount of computation that scales exponentially with the combination size and in most cases is intractable even for cutting-edge supercomputers. We propose a parameter-free heuristic approach that leverages the intrinsic topology of gene-patient mutations to identify carcinogenic combinations. The biological relevance of the identified combinations is measured by using them to predict the presence of tumor in previously unseen samples. The resulting classifiers for 16 cancer types perform on par with exhaustive search results, and score the average of 80.1% sensitivity and 91.6% specificity for the best choice of hit range per cancer type. Our approach is able to find higher-hit carcinogenic combinations targeting which would take years of computations using exhaustive search.

The earth sciences research community has an unprecedented opportunity to exploit the vast amount of data available from earth observation (EO) satellites and earth system models (ESM). The ascent and application of artificial intelligence foundation models (FM) can be attributed to the availability of large volumes of curated data, access to extensive computing resources and the maturity of deep learning techniques. Vision transformers (ViT) architectures have been adapted for image and image-like data, such as EO data and ESM simulation output. Pretraining foundation models is a compute intensive process, often requiring 105- 107GPU hours for large scale scientific applications. There is a limited body of knowledge on compute optimal methods for pretraining, necessitating a trial and error process. We have performed a series of experiments using ViT backbones at different scales to understand optimal and cost-effective ways to improve scientific throughput. This preliminary benchmark provides an assessment of which architectures and model configurations are favorable in a given scientific context.

Cancer is a leading cause of death in the US, and it results from a combination of two-nine genetic mutations. Identifying five-hit combinations responsible for several cancer types is computationally intractable even with the fastest super-computers in the USA. Iterating through nested loops required by the process presents a simplex-shaped workload with irregular memory access patterns. Distributing this workload efficiently across thousands of GPUs offers a challenge in dividing simplex-shaped (triangular/tetrahedral) workload into similar shapes with equal volume. Irregular memory access patterns create imbalanced compute utilization across nodes. We developed a generalized solution for distributing a simplex-shaped workload by partially coalescing the nested for-loops, minimizing the memory access overhead by efficiently utilizing limited shared memory, a dynamic scheduler, and loop tiling. For 4-hit combinations, we achieved a 90% − 100% strong scaling efficiency for up to 3594 V100 GPUs on the Summit supercomputer. Finally, we designed and implemented a distributed algorithm to identify 5-hit combinations for four different cancer types, and the identified combinations can differentiate between cancer and normal samples with 86.59−88.79% precision and 84.42 − 90.91% recall. We also demonstrated the robustness of our solution by porting the code to another leadership class computing platform Crusher, a testbed for the fastest supercomputer Frontier. On Crusher, we achieved 98% strong scaling efficiency on 50 nodes (400 AMD MI250X GCDs) and demonstrated the computational readiness of Frontier for scientific applications.

Large language models (LLMs) are poised to revolutionize the way we conduct scientific research. However, both model complexity and pre-training cost are impeding effective adoption for the wider science community. Identifying suitable scientific use cases, finding the optimal balance between model and data sizes, and scaling up model training are among the most pressing issues that need to be addressed. In this study, we provide practical solutions for building and using LLM-based foundation models targeting scientific research use cases. We present an end-to-end examination of the effectiveness of LLMs in scientific research, including their scaling behavior and computational requirements on Frontier, the first Exascale supercomputer. We have also developed for release to the scientific community a suite of open foundation models called FORGE with up to 26B parameters using 257B tokens from over 200M scientific articles, with performance either on par or superior to other state-of-the-art comparable models. We have demonstrated the use and effectiveness of FORGE on scientific downstream tasks. Our research establishes best practices that can be applied across various fields to take advantage of LLMs for scientific discovery.

The peculiar nature of whole slide imaging (WSI), digitizing conventional glass slides to obtain multiple high resolution images which capture microscopic details of a patient’s histopathological features, has garnered increased interest from the computer vision research community over the last two decades. Given the unique computational space and time complexity inherent to gigapixel-size whole slide image data, researchers have proposed novel machine learning algorithms to aid in the performance of diagnostic tasks in clinical pathology. One effective algorithm represents a Whole slide image as a bag of smaller image patches, which can be represented as low-dimension image patch embeddings. Weakly supervised deep-learning methods, such as cluster-constrained-attention multiple instance learning (CLAM), have shown promising results when combined with image patch embeddings. While traditional ensemble classifiers yield improved task performance, such methods come with a steep cost in model complexity. Through knowledge distillation, it is possible to retain some performance improvements from an ensemble, while minimizing costs to model complexity. In this work, we implement a weakly supervised ensemble using clustering-constrained-attention multiple-instance learners (CLAM), which uses attention and instance-level clustering to identify task salient regions and feature extraction in whole slides. By applying logit-based and attention-based knowledge distillation, we show it is possible to retain some performance improvements resulting from the ensemble at zero cost to model complexity.

Scientific communities are increasingly adopting deep learning (DL) models in their applications to accelerate scientific discovery processes. However, with rapid growth in the computing capabilities of HPC supercomputers, large-scale DL applications have to spend a significant portion of training time performing I/O to a parallel storage system. Previous research works have investigated optimization techniques such as prefetching and caching. Unfortunately, there exist non-trivial challenges to adopting the existing solutions on HPC supercomputers for large-scale DL training applications, which include non-performance and/or failures at extreme scale, lack of portability and generality in design, complex deployment methodology, and being limited to a specific application or dataset. To address these challenges, we propose High-Velocity AI Cache (HVAC), a distributed read-cache layer that targets and fully exploits the node-local storage or near node-local storage technology. HVAC seamlessly accelerates read I/O by aggregating node-local or near node-local storage, avoiding metadata lookups and file locking while preserving portability in the application code. We deploy and evaluate HVAC on 1,024 nodes (with over 6000 NVIDIA V100 GPUS) of the Summit supercomputer. In particular, we evaluate the scalability, efficiency, accuracy, and load distribution of HVAC compared to GPFS and XFS-on-NVMe. With four different DL applications, we observe an average 25 % performance improvement atop GPFS and 9% drop against XFS-on-NVMe, which scale linearly and are considered the performance upper bound. We envision HVAC as an important caching library for upcoming HPC supercomputers such as Frontier.

Cancer is a leading cause of death in the US, second only to heart disease. It is primarily a result of a combination of an estimated two-nine genetic mutations (multi-hit combinations). Although a body of research has identified hundreds of cancer-causing genetic mutations, we don't know the specific combination of mutations responsible for specific instances of cancer for most cancer types. An approximate algorithm for solving the weighted set cover problem was previously adapted to identify combinations of genes with mutations that may be responsible for individual instances of cancer. However, the algorithm's computational requirement scales exponentially with the number of genes, making it impractical for identifying more than three-hit combinations, even after the algorithm was parallelized and scaled up to a V100 GPU. Since most cancers have been estimated to require more than three hits, we scaled out the algorithm to identify combinations of four or more hits using 1000 nodes (6000 V100 GPUs with ≈ 48×106processing cores) on the Summit supercomputer at Oak Ridge National Laboratory. Efficiently scaling out the algorithm required a series of algorithmic innovations and optimizations for balancing an exponentially divergent workload across processors and for minimizing memory latency and inter-node communication. We achieved an average strong scaling efficiency of 90.14% (80.96%-97.96% for 200 to 1000 nodes), compared to a 100 node run, with 84.18% scaling efficiency for 1000 nodes. With experimental validation, the multi-hit combinations identified here could provide further insight into the etiology of different cancer subtypes and provide a rational basis for targeted combination therapy.