MRI: Acquisition of an Extreme GPU Cluster for Interdisciplinary Research

斯坦福大学要求在36个月内获得350万美元的超级GPU高性能计算集群，称为X-GPU，包括使用Cray Hydra技术和FDR Infiniband构建的54个计算节点。每个节点都有Intel Haswell 12核；8张NVIDIA Kepler显卡；128gb DDR4内存；一个120gb的SSD和两个1tb的硬盘。高效节能的计算设备，提供近千万亿次的计算能力。它将被1)至少25个研究小组使用，代表斯坦福大学15个系和4个学院的100多名学生和博士后；2)来自全国至少7个其他机构的至少8名合作者；3)通过nsf赞助的XSEDE分配系统，由多达数百名国家研究人员使用。pi计划向XSEDE提供25%的X-GPU，以抵消Keeneland计划退役的影响，目前XSEDE资源为全国社区提供cpu和gpu的异构并行计算。通过该仪器确定的科学成果包括但不限于：天体物理学和宇宙学、生物信息学和生物学、材料建模和气候建模。研究人员已经投入了大量的精力来开发建模和仿真代码，这些代码可以在gpu加速的集群上展示高性能。pi计划开发软件基础设施和教育材料，以帮助国家社区过渡到细粒度并行思维和算法设计，这对于有效使用这种新型高性能、低成本、节能的架构至关重要。

Stanford University requests $3,500,000 over 36 months to acquire an extreme GPU HPC cluster, called X-GPU, comprising 54 compute nodes built using the Cray Hydra technology with FDR Infiniband. Each node has Intel Haswell 12-cores; 8 NVIDIA Kepler cards; 128 GB of DDR4 memory; a 120 GB SSD and two 1 TB hard drives. energy-efficient, computational facility providing almost a petaflop of computational power. It will be used by 1) at least 25 research groups representing more than 100 students and postdoctorals at Stanford across 15 departments and 4 schools, 2) at least 8 collaborators from at least 7 other institutions across the nation, and 3) by as many as hundreds of national researchers through the NSF-sponsored XSEDE allocation system. The PIs plan to offer 25% of X-GPU to XSEDE to offset the impacts from the planned retiring of Keeneland, the current XSEDE resource providing heterogeneous parallel computing with CPUs and GPUs to the national community.Identified scientific outcomes enabled by this instrument include, but not limited to: astrophysics and cosmology, bioinformatics and biology, materials modeling, and climate modeling. The researchers have already invested significant efforts to develop modeling and simulation codes that can demonstrate high performance on GPU-accelerated clusters. The PIs plan develop software infrastructure and educational materials to help the national community in the transition to fine-grained parallel thinking and algorithm design, which is critical to effectively use this novel high-performance, low-cost, energy-efficient architecture.