UBC ML Computational Cluster

机器学习研究在很大程度上依赖于计算资源的可用性。不幸的是，由于高需求，并不是所有的教职员工都能够获得现有的高性能计算设施的分配。此外，这些系统上的分配，虽然适合于广泛的计算，但并不是为达到最终模型结构所需的开发工作而设计的。最后，这些设施中可用的硬件不一定最适合机器学习应用程序，因为这些设施旨在为所有学科提供服务。 为了解决这些问题，我们建议建立一个可扩展的本地计算集群，配备最新的NVIDIA RTX-3090和100 TB的存储，以支持5名教师和50多名计算机科学和统计学研究生和博士后研究员进行的研究，他们的研究由NSERC，CIFAR，MITACS和其他来源资助。该系统旨在补充现有高性能计算设施的分配，如加拿大计算和UBC ARC Sockeye。该系统的设计是可扩展的，因此如果需要，可以添加新的硬件，新的教师可以在未来几年访问。

Machine learning research relies heavily on the availability of computational resources. Unfortunately, due to high demand, not all faculty members are able to gain access to allocations on existing high performance computing facilities. Moreover, allocations on these systems, while suitable for extensive computation, are not designed for developmental work needed to reach a final model structure. Finally, the hardware available at these facilities is not necessarily most suited to machine learning applications, as the facilities are designed to service all disciplines. To resolve these issues, we propose setting up an extendable local computational cluster, equipped with the latest NVIDIA RTX-3090 and 100 TB of storage to support research performed by 5 faculty members and more than 50 graduate students and postdoctoral fellows in computer science and statistics, whose research is funded by NSERC, CIFAR, MITACS and other sources. This system is designed to complement allocations from existing high performance computing facilities such as Compute Canada and UBC ARC Sockeye. The system is designed to be extendable, so new hardware can be added if needed, and new faculty members can be given access in future years.