SEARCH · Search NASA
Results for “high-bandwidth memory”
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Runtime extension for neural network training with heterogeneous memory
Systems, apparatuses, and methods for managing buffers in a neural network implementation with heterogeneous memory are disclosed. A system includes a neural network coupled to a first memory and a second memory. The first memory is a relatively low-capacity, high-bandwidth memory while the second memory is a relatively high-capacity, low-bandwidth memory. During a forward propagation pass of the neural network, a run-time manager monitors the usage of the buffers for the various layers of the neural network. During a backward propagation pass of the neural network, the run-time manager determines how to move the buffers between the first and second memories based on the monitored buffer usage during the forward propagation pass. As a result, the run-time manager is able to reduce memory access latency for the layers of the neural network during the backward propagation pass.
Pushing the limits of NAND technology scaling with ferroelectrics
Artificial intelligence (AI) continues to drive transformative advancements across various industries. The data-intensive nature of AI training (and inferencing) has resulted in the generation of unprecedented volumes of data with machine-generated content surpassing human-generated data by more than 100-fold in 2025. Efficiently managing this data influx necessitates advanced digital storage technologies. However, traditional NAND flash memory, which is critical for supporting data flows in AI systems—alongside high-bandwidth memory, for AI training—faces fundamental scaling limitations as it approaches the 1000-layer milestone, encompassing more than 40 trillion transistors. This article delves into the potential of hafnia-based ferroelectric materials as a breakthrough solution to these challenges. Recent advancements indicate that the intrinsic limitations of ferroelectric field-effect transistors (FEFETs) can be mitigated through material and device-level engineering. These advancements enable FEFETs to meet the stringent density, reliability, and scalability requirements of future three-dimensional NAND technology. The role of ferroelectrics in addressing NAND scaling challenges and expanding storage capabilities presents a promising avenue for meeting the storage demands of the AI-driven era.
Performance Potential of Mixed Data Management Modes for Heterogeneous Memory Systems
Many high-performance systems now include different types of memory devices within the same compute platform to meet strict performance and cost constraints. Such heterogeneous memory systems often include an upper-level tier with better performance, but limited capacity, and lower-level tiers with higher capacity, but less bandwidth and longer latencies for reads and writes. To utilize the different memory layers efficiently, current systems rely on hardware-directed, memory -side caching or they provide facilities in the operating system (OS) that allow applications to make their own data-tier assignments. Since these data management options each come with their own set of trade-offs, many systems also include mixed data management configurations that allow applications to employ hardware- and software-directed management simultaneously, but for different portions of their address space. Despite the opportunity to address limitations of stand-alone data management options, such mixed management modes are under-utilized in practice, and have not been evaluated in prior studies of complex memory hardware. In this work, we develop custom program profiling, configurations, and policies to study the potential of mixed data management modes to outperform hardware- or software-based management schemes alone. Our experiments, conducted on an Intel ® Knights Landing platform with high-bandwidth memory, demonstrate that the mixed data management mode achieves the same or better performance than the best stand-alone option for five memory intensive benchmark applications (run separately and in isolation), resulting in an average speedup compared to the best stand-alone policy of over 10 %, on average.
Understanding GPU Memory Corruption at Extreme Scale: The Summit Case Study
GPU memory corruption and in particular double-bit errors (DBEs) remain one of the least understood aspects of HPC system reliability. Albeit rare, their occurrences always lead to job termination and can potentially cost thousands of node-hours, either from wasted computations or as the overhead from regular checkpointing needed to minimize the losses. As supercomputers and their components simultaneously grow in scale, density, failure rates, and environmental footprint, the efficiency of HPC operations becomes both an imperative and a challenge. We examine DBEs using system telemetry data and logs collected from the Summit supercomputer, equipped with 27,648 Tesla V100 GPUs with 2nd-generation high-bandwidth memory (HBM2). Using exploratory data analysis and statistical learning, we extract several insights about memory reliability in such GPUs. We find that GPUs with prior DBE occurrences are prone to experience them again due to otherwise harmless factors, correlate this phenomenon with GPU placement, and suggest manufacturing variability as a factor. On the general population of GPUs, we link DBEs to short- and long-term high power consumption modes while finding no significant correlation with higher temperatures. We also show that the workload type can be a factor in memory’s propensity to corruption.
OLCF Summit Supercomputer GPU Snapshots During Double-Bit Errors and Normal Operations
As we move into the exascale era, the power and energy footprints of high-performance computing (HPC) systems have grown significantly larger. Due to the harsh power and thermal conditions the system, components are exposed to extreme operating conditions. Operation of such modern HPC systems requires deep insights into long term system behavior to maintain its efficiency as well as its longevity. To help the HPC community to gain such insights, we provide double-bit errors using system telemetry data and logs collected from the Summit supercomputer, equipped with 27,648 Tesla V100 GPUs with 2nd-generation high-bandwidth memory (HBM2). The dataset relies on Nvidia XID records internally collected by GPU firmware at the time of failure occurrence, on the reboot-time logs of each Summit node, on node-level job scheduler records collected after each job termination, and on a 1Hz data rate from the baseboard management controllers (BMCs) of each Summit compute node using the OpenBMC event subscription protocol. Technical details can be found in the paper Oles et. al “Understanding GPU Memory Corruption at Extreme Scale: The Summit Case Study” ICS’24 (https://doi.org/10.1145/3650200.3656615).
JACC.shared: Leveraging HPC Metaprogramming and Performance Portability for Computations That Use Shared Memory GPUs
In this work, we present JACC.shared, a new feature of Julia for ACCelerators (JACC), which is the performanceportable and metaprogramming model of the just-in-time and LLVM-based Julia language. This new feature allows JACC applications to leverage the high-performance computing (HPC) capabilities of high-bandwidth, on-chip GPU memory. Historically, exploiting high-bandwidth, shared-memory GPUs has not been a priority for high-level programming solutions. JACC.shared covers that gap for the first time, thereby providing a highlevel, portable, and easy-to-use solution for programmers to exploit this memory and supporting all current major accelerator architectures. Well-known HPC and AI workloads, such as multi/hyperspectral imaging and AI convolutions, have been used to evaluate JACC.shared on two exascale GPU architectures hosted by some of the most powerful US Department of Energy supercomputers: Perlmutter (NVIDIA A100) and Frontier (AMD MI250X). The performance evaluation reports speedup of up to 3.5× by adding only one line of code to the base codes, thus providing important accelerators in a simple, portable, and transparent way and elevating the programming productivity and performance-portability capabilities for Julia/JACC HPC, AI, and scientific applications.
The ECP SICM project: Managing complex memory hierarchies for exascale applications
The Exascale Computing Project (ECP)’s Simplified Interface to Complex Memories (SICM) effort focuses on developing universal interfaces for discovering, managing, and sharing data across complex memory hierarchies. These facilitate the exploitation of emerging memory technologies and support precise control over their various trade-offs such as high-bandwidth versus low-latency, persistent versus ephemeral, high-capacity versus low-capacity, and near-CPU versus near-GPU. SICM comprises three interrelated components: a low-level interface, a high-level interface, and a persistent-heap interface. The low-level SICM interface is intended for system and run-time developers as well as expert application developers who prefer full control of the memory objects used within their application. The high-level SICM interface builds upon the low-level interface, employing application-level profiling and analysis to optimize data management for complex memory hierarchies. The persistent-heap interface provides applications with a persistent memory allocator that can allocate custom C++ data structures in both block-storage and byte-addressable persistent memories.
Multiplexed color centers in a silicon photonic cavity array
Entanglement distribution is central to the modular scaling of quantum processors and establishing quantum networks. Color centers with telecom-band transitions and long spin coherence times are suitable candidates for long-distance entanglement distribution. However, high-bandwidth memory-enhanced quantum communication is limited by high-yield, scalable creation of efficient spin-photon interfaces. Here, we develop a silicon photonics platform consisting of arrays of bus-coupled cavities. The coupling to a common bus waveguide enables simultaneous access to individually addressable cavity-enhanced T center arrays. We demonstrate frequency-multiplexed operation of two T centers in separate photonic crystal cavities. In addition, we investigate the cavity enhancement of a T center through hybridized modes formed between physically distant cavities. Our results show that bus-coupled arrays of cavity-enhanced color centers could enable efficient on-chip and long-distance entanglement distribution.
MOSIQS: Persistent Memory Object Storage With Metadata Indexing and Querying for Scientific Computing
Scientific applications often require high-bandwidth shared storage to perform joint simulations and collaborative data analytics. Shared memory pools provide a chance to satisfy such needs. Recently, a high-speed network such as Gen-Z utilizing persistent memory (PM) offers an opportunity to create a shared memory pool connected to compute nodes. However, there are several challenges to use scientific applications on the shared memory pool directly such as scalability, failure-atomicity, and lack of scientific metadata-based search and query. In this paper, we propose MOSIQS, a persistent memory object storage framework with metadata indexing and querying for scientific computing. We design MOSIQS based on the key idea that memory objects on PM pool can live beyond the application lifetime and can become the sharing currency for applications and scientists. MOSIQS provides an aggregate memory pool atop an array of persistent memory devices to store and access memory objects to accelerate scientific computing. MOSIQS uses a lightweight persistent memory key-value store to manage the metadata of memory objects, which enables memory object sharing. To facilitate metadata search and query over millions of memory objects resident on memory pool, we introduce Group Split and Merge (GSM), a novel persistent index data structure designed primarily for scientific datasets. GSM splits and merges dynamically to minimize the query search space and maintains low query processing time while overcoming the index storage overhead. MOSIQS is implemented on top of PMDK. We evaluate the proposed approach on many-core server with an array of real PM devices. Experimental results show that MOSIQS gains a 100% write performance improvement and executes multi-attribute queries efficiently with 2.7× less index storage overhead offering significant potential to speed up scientific computing applications.
Optically connected memory for disaggregated data centers
Recent advances in integrated photonics enable the implementation of reconfigurable, high-bandwidth, and low energy-per-bit interconnects in next-generation data centers. We propose and evaluate an Optically Connected Memory (OCM) architecture that disaggregates the main memory from the computation nodes in data centers. OCM is based on micro-ring resonators (MRRs), and it does not require any modification to the DRAM memory modules. We calculate energy consumption from real photonic devices and integrate them into a system simulator to evaluate performance. Here, our results show that (1) OCM is capable of interconnecting four DDR4 memory channels to a computing node using two fibers with 1.02 pJ energy-per-bit consumption and (2) OCM performs up to 5.5× faster than a disaggregated memory with 40G PCIe NIC connectors to computing nodes.
nng_stream
This project uses the NanoMsg next-generation library to receive data, temporarily store in an in-memory buffer, and then send it out. The buffer handles a high-bandwidth stream from multiple producers (sources) to multiple consumers (sinks).
Deployment and Evaluation of SciStream on OLCF's Advanced Computing Ecosystem (ACE)
The growing demand for real-time analysis, experimental steering, and decision-making in scientific workflows has created a need for tightly coupled integrations between experimental facilities and high-performance computing (HPC) systems. The Department of Energy’s Integrated Research Infrastructure (IRI) initiative highlights data streaming as a key capability for enabling memory-to-memory data transfers, bypassing the limitations of traditional store-and-forward models. SciStream is a toolkit developed by researchers at Argonne National Laboratory (ANL) to support such streaming by addressing cross-domain security, delegated authentication, and application transparency. We deployed and evaluated SciStream on the Oak Ridge Leadership Computing Facility’s (OLCF) Advanced Computing Ecosystem (ACE) infrastructure, leveraging the Olivine OpenShift cluster and its high-bandwidth Data Streaming Nodes (DSNs) as gateway nodes. Our evaluation included synthetic streaming workloads derived from IRI science workflows, a streaming simulator, and integration with RabbitMQ to handle low-level messaging. This report documents the deployment process, performance evaluation, and challenges encountered, along with opportunities for future improvements.
Multimode Strong Coupling in Cavity Optomechanics
Optomechanical systems show great potential as quantum transducers and information storage devices for use in future hybrid quantum networks. In this context, optomechanical strong coupling can enable efficient, high-bandwidth, and deterministic transfer of quantum states. While optomechanical strong coupling has been realized at optical frequencies, it has proven difficult to identify a robust optomechanical system that features the low loss and high coupling rates required for more sophisticated control of mechanical motion. In this paper, we demonstrate strong coupling in a Brillouin-based bulk cavity optomechanical system in both the single-mode and the multimode strong-coupling regime, which leads to a useful device both for applications in quantum information and for investigating decoherence phenomena in bulk acoustic wave resonators. Using nontrivial mode hybridizations in the strong-coupling regime, we create hybridized photonic-phononic modes with lifetimes that are significantly longer than those of the uncoupled system. This surprising lifetime enhancement, which results from the interference of decay channels, showcases the use of multimode strong coupling as a general strategy to control extrinsic decoherence mechanisms. Moreover, phonons supported by such bulk-acoustic-wave resonators have a collection of properties, including high frequencies, long coherence times, and robustness against thermal decoherence, that make this optomechanical system particularly enticing for applications such as quantum transduction and memories. Hence, this system provides access to phenomena in a previously unexplored regime of optomechanical interactions and could serve as an important building block for future quantum devices.
IceNet for FireBox - A Berkeley Warehouse-Scale Computer
Berkeley’s FireBox is a next-generation warehouse-scale computer (WSC) that utilizes the energy-efficiency and bandwidth density of integrated silicon-photonic interconnects to enable a new high-bandwidth and low-latency network fabric connecting thousands of compute nodes to petabytes of DRAM and Flash storage. The high bandwidth, low latency and high connectivity of FireBox’s WSC network fabric (IceNet) will enable dramatic improvements in the overall system energy efficiency enabling fine-grain power control on system resources (processors, links and memory/storage components). IceNet is a special 3-stage photonic Clos network architected to achieve ultra-low-latency connectivity between processor nodes and memory, drastically cutting down on the energy wasted in resource idling (processors and memory stalled due to pending network requests). This is achieved by integration of the first and last switch stages into processor/memory hub clients and by heavy over-provisioning of the high-radix middle switches (FlareSwitches). A key hardware component developed in this program is an active laser power management photonic integrated circuits called LightSpark. It interacts with the FlareSwitch and provides laser power to a subset of occupied switch ports, increasing the utilization of laser light in the photonic network by an order of magnitude. In addition to guiding the laser power where it is needed, the laser-power management module enables both wavelength and laser redundancy, significantly increasing the robustness of the system. The goal of the IceNet fabric is to enable communication between 1000s of processor nodes and PBs of memory/storage with <100ns latency, <10pJ/b wall-plug energy cost at multiple Pb/s of available connectivity bandwidth. These metrics represent two-orders of magnitude improvement with respect to the status of current data-center technology.