Request for Dell PowerEdge XE9680 8x H100 NVIDIA GPU server

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

This request is for funds to purchase a Dell PowerEdge XE9680 8x H100 GPU server with NV-Link. The instrument will be housed within the Center for Biologic Imaging (CBI) a large, centralized imaging facility/Center serving the needs of NIH supported researchers across our campus and at institutions both local and across the US. A focus of this research Center is to develop, implement and integrate both traditional and new optical imaging approaches into modern biomedical research. Data reconstruction and quantitation is an essential arm of our research product and is integrated into experimental protocols from first conception. Over the last 6 years we have been increasingly involved in the development of imaging platforms and development of data processing workflows that require significant compute resources, primarily GPU, which far exceed the processing capabilities of desktop workstations. The center currently houses a dedicated machine room with several petabytes of storage and HPC cluster with 2 NVIDIA K80 GPU nodes. However, data acquisition has rapidly outpaced our ability to process it using the current infrastructure. The biggest data producers for the Center are; 1: Ribbon scanning confocal, 2: light sheet tools for large tissue samples (MesoSPIM) and 3: multi-photon microscopy. The data sets generated by these methods are commonly between 10 and 100 terabytes in size and cannot be rationally processed with conventional computing solutions (single GPU workstations) and truly require modern computational architecture. The device requested here, a 8x H100 GPU server with NV-Link, is configured to meet these needs. To date, we have used shared HPC resources to supplement our computational needs, but these resources are neither intended for nor appropriate to use for our high-demand environment. At this point, we feel that there is a very compelling argument to build rationally scaled local capability. The fundamental reasons are difficulty of data movement and the absolute variety and scale of the projects we conduct in our center being outside the interests of these shared resources. The scientific diet of the Center ranges from Tissue Organoids to Pulmonary Medicine, from Ocular Biology to Tumor Biology or expanded tissues or entire organisms (such as zebra fish). Because of this it is very difficult to define and describe the diversity of research projects that this infrastructure will be used for. Accordingly, we have limited our application to several “major user” projects which produce large datasets from high-speed imaging systems and have an outsized need for GPU resources. Research areas include Neuroscience, Pulmonary Medicine, Viral Pathogenesis, Collagen in health and disease, and Imaging Technology development. Each of these major users needs to contextualize molecular or cellular events in the context of the whole organ or organisms at cellular resolution or better. As this involves highly anisotropic tissues, or complex/rare interactions, the questions cannot be addressed outside the context of the whole organ. Each illustrates a clear need for the compute capability/platform we are requesting.