Category II : Ookami: A high-productivity path to frontiers of scientific discovery enabled by exascale system technologies

纽约的州立大学提议采购日本以外的第一台计算机，并将其运行至少四年，该计算机采用富士通为日本的艾级计算之路开发的A64 fx处理器（即，每秒10^18次运算的计算机）。这款基于ARM的多核512位SIMD向量处理器具有超高带宽存储器，有望保留熟悉且成功的编程模型，同时为包括仿真和大数据在内的广泛应用实现极高的性能。 该测试平台极大地扩展了当前NSF赞助的HPC技术，并将使社区能够评估和展示该技术在多种环境中部署的潜力。通过与NSF的极限科学和工程发现环境（XSEDE）集成，该系统将被广泛访问，并充分利用现有的网络基础设施，包括XDMoD监控系统。这对科学意味着什么？与部署期间预期的最佳CPU相比，A64 fx在内存密集型应用程序（如许多工程和物理代码中的稀疏矩阵求解器）上的性能提高了2- 4倍。对于几乎所有其他应用程序，性能也更好或具有竞争力。这种转换性能应该几乎是开箱即用的，通过调优可以获得额外的性能。对于科学家或工程师来说，这意味着更快的解决方案，大大减少了程序员的工作量。要启用的目标应用程序类别是内存带宽密集型应用程序，每个节点内存为32 GB，预计许多其他应用程序将获得显著收益。对XSEDE工作负载的分析表明，86%的作业（85%的周期）将适合每个节点的可用内存，并且大多数作业是内存带宽密集型的。最后，我们有具体的计划，通过与机构、地区和国家层面的外部组织合作，大幅扩大对科学和工程研究的参与。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The State University of New York proposes to procure and operate for at least four years the first computer outside of Japan with the A64fx processor developed by Fujitsu for the Japanese path to exascale computing (i.e., computers capable of 10^18 operations per second). The ARM-based, multi-core, 512-bit SIMD-vector processor with ultrahigh-bandwidth memory promises to retain familiar and successful programming models while achieving very high performance for a wide range of applications including simulation and big data. The testbed significantly extends current NSF-sponsored HPC technologies and will enable the community to evaluate and demonstrate the potential of this technology for deployment in multiple settings. Through integration with NSF's Extreme Science and Engineering Discovery Environment (XSEDE), the system will be widely accessible and fully leverages existing cyber infrastructure including the XDMoD monitoring system.What does this mean for science? Compared with the best CPUs anticipated during the deployment period, A64fx offers 2-4x better performance on memory-intensive applications such as sparse-matrix solvers found in many engineering and physics codes. For nearly all other applications, performance is also better or competitive. This transformational performance should be available nearly out of the box, with additional performance possible from tuning. To the scientist or engineer this means faster time to solution with significantly less programmer effort. The target class of applications to be enabled are memory-bandwidth intensive with 32GB/node memory, with significant gains anticipated for many other applications. Analysis of XSEDE workload shows 86% of all jobs (85% cycles) will fit within the available memory per node and that the majority of jobs are memory-bandwidth intensive. Finally, we have concrete plans to substantially broaden participation in science and engineering research by partnering with external organizations at the institutional, regional, and national levels.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.