Jan 2026· IEEE Access· Vol 14, pp. 156028-156042· 1 citation· 46 references
Computer Science
TL;DR
This paper evaluates four platform–mechanism combinations across a datacenter-class GPU and an edge computing platform, evaluating the capability of these mechanisms to guarantee bounded execution times and freedom from interference across both classical neural network architectures and more computationally demanding models such as Vision Transformers.
Abstract
Traditional safety-critical architectures distribute software functionalities across separate hardware components. However, the increasing computational demands of modern applications make such federated approaches difficult to scale, motivating the need for efficient software-based resource sharing on powerful accelerators. In this paper, we investigate how existing features of modern GPUs can be leveraged to satisfy the isolation and performance requirements of safety-critical applications while maintaining high resource utilization. We analyze the main mechanisms that enable concurrent GPU usage across processes, ranging from lightweight context-switching approaches, such as Multi-Process Service (MPS), to strict resource-partitioning strategies, such as Multi-Instance GPU (MIG) and Green Contexts (GC). Our analysis covers both general-purpose GPUs and devices specifically designed for edge computing environments. We evaluate four platform–mechanism combinations across a datacenter-class GPU (NVIDIA A100) and an edge computing platform (NVIDIA Jetson AGX Orin), analyzing the capability of these mechanisms to guarantee bounded execution times and freedom from interference, which are key requirements in safety-related systems, across both classical neural network architectures and more computationally demanding models such as Vision Transformers (ViTs).
GPUs are expensive resources but often underutilized. This inefficiency is exacerbated by the increasing diversity of AI workloads, where small or specialized models often fail to saturate high-capacity devices, necessitating GPU sharing. Existing sharing approaches force a hard tradeoff. They either partition hardware...
Yi-Cheng Liu, Yi-Fan Qiao, Tian Xia et al.· Proceedings of the 32nd ACM...· 0 citations
This work investigates the memory capabilities of the NVIDIA DGX Spark, a novel platform featuring a unified memory architecture where DDR memory is located on the CPU and is fully accessible from the GPU.
Silvia R. Alcaraz, S. Hepkema, Vasilis Mageirakos et al.· DIMES@SOSP· 0 citations
Ghost is an OS-level GPU virtualization layer integrated directly into the open-source GPU driver, using a GPU container abstraction with cgroup -like APIs for compute and memory control and privileged hardware-level scheduling and preemption for dynamic compute resource management.
Exo-GPU, an imperative, low-level language that creates minimal abstraction over CUDA, is proposed, to treat parallelism and synchronization as mere annotations on sequential code rather than as fundamental control flow primitives, enabling verification that these constructs do not alter the program semantics.
David Akeley, Yuka Ikarashi, Jonathan Ragan-Kelley· arXiv.org· 2 citations
NVIDIA's Grace Hopper (GH200) superchip represents a new generation of heterogeneous computing, tightly integrating a 72-core Arm-based CPU with a Hopper GPU on a single module via a high-bandwidth, cache-coherent interconnect. By eliminating the conventional PCIe bottleneck for CPU-GPU data movement, this design estab...
Ravan Nazaraliyev, Yi-Cheng Zhang, N. Abu-Ghazaleh· 2 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.