Graphcore IPU benchmarking and development

Graphcore IPU 基准测试和开发

• 批准号: ST/W006871/1
• 负责人: Owain Kenway
• 金额: $ 17.53万
• 依托单位: University College London
• 依托单位国家: 英国
• 项目类别: Research Grant
• 财政年份: 2021
• 资助国家: 英国
• 起止时间: 2021 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=ST%2FW006871%2F1
• 关键词: Graphcore; IPU; benchmarking; development

In 2018, the Exascale Computing ALgorithms & Infrastructures for the Benefit of UK Research (ExCALIBUR) programme was proposed by the Met Office, CCFE and EPSRC (on behalf of UKRI). The goal of ExCALIBUR is to redesign high priority computer codes and algorithms, keeping UK research and development at the forefront of high-performance simulation science. The challenge spans many disciplines and as such the programme of research will be delivered through a partnership between the Met Office and UKRI Research Councils. Research software engineers and scientists will work together to future proof the UK against the fast-moving changes in supercomputer designs. This combined scientific expertise will push the boundaries of science across a wide range of fields delivering transformational change at the cutting-edge of scientific supercomputing. DiRAC proposed the inclusion in the ExCALIBUR business case of a request for £4.5M in capital funding over 4.5 years to develop a hardware fore-sighting programme. Industry co-funding for the programme will be sought where possible. The £4.5m capital is intended to provide a testbed area that uses pre-commercial equipment for software prototyping and development. It has two main purposes: (1) to enable the software community to be ready to use commercial products effectively as soon as they come on to the market; and (2) to provide the UKRI HPC community with the ability to influence industry and the necessary knowledge to guide their purchase decisions. This will ensure that facilities and the future UK National e-Infrastructure are in a position to maximise value for money by getting the most powerful systems exactly suited to the communities' needs. This double-pronged approach will give UK researchers a competitive advantage internationally. ExCALIBUR will now establish a set of modest-sized, adaptable clusters dedicated solely to this purpose and embedded within established HPC environments. Although small, they need to be of a scale capable of carrying out meaningful performance studies. They are expected to be co-funded with industry partners and will initially require investments of £200k-£300k each, and will allow a range of future hardware to be assessed for its relevance to the delivery of UKRI science and innovation. The pre-commercial equipment will be refreshed and added to on a regular, likely to be annual, basis. This agile tactic is designed to take advantage of the different approaches across industry (some companies, e.g. NVidia tend to have a short (less than 3-month) pre-commercial window while for others this can be up to a year). ExCALIBUR can use the hardware piloting systems to drive software innovation across the UKRI research community. Researchers are rightly reluctant to invest time in code development to take advantage of new hardware which may not be available at scale for several years or may even prove not to have longevity - scientific leadership demands that research funding is used to deliver science results now. In addition and DiRAC and others will offer funded RSE effort to support the development work combined with access to novel technologies within modest-sized systems, Excalibur can lower the bar for engaging with the process of software re-engineering and encourage researchers to make the necessary (modest) investments of their time. In some cases, there may also be the potential for some immediate science outputs by exploiting the proof-of-concept systems. Excalibur will thus be able to provide an incentive for greater software innovation across the UKRI research communities and help to ensure that when novel technology is included in national services, there are workflows that are already able to exploit it optimally. This will increase productivity across all UKRI computing services and enable UK researchers to use the latest hardware to deliver the largest and most complex calculations, ensuring international leadership.

2018年，MET办公室，CCFE和EPSRC（代表Ukri）提出了针对英国研究（Excalibur）计划的Exascale计算算法和基础架构（Excalibur）计划。 Excalibur的目标是重新设计高优先级计算机代码和算法，使英国的研发位于高性能模拟的最前沿。挑战涵盖了许多学科，因此研究计划将通过大都会办公室和乌克里研究委员会之间的合作伙伴关系实现。研究软件工程师和科学家将共同努力，以证明英国不受超级计算机设计的快速变化。这位结合的科学专家将在科学超级计算的最先进的情况下推动各种领域的科学界限。狄拉克（Dirac）提议将其包含在Excalibur业务案例中，要求在4。5年内提供450万英镑的资本资金，以制定硬件前视计划。在可能的情况下，该计划的行业将是有意义的。 450万英镑的资本旨在提供经过商业前设备进行软件原型制作和开发的测试区域。它有两个主要目的：（1）使软件社区能够在进入市场后立即有效地使用商业产品； （2）为UKRI HPC社区提供影响行业的能力以及指导其购买决策的必要知识。这将确保设施和未来的英国国家电子基础设施可以通过获得完全适合社区需求的最强大的系统来最大程度地价值。这种双管责任的方法将使英国研究人员在国际上具有竞争优势。 Excalibur现在将建立一组适度的，适应性的群集，专门用于此目的，并嵌入在已建立的HPC环境中。尽管很小，但它们需要具有能够进行有意义的绩效研究的规模。预计他们将与行业合作伙伴共同资助，最初需要每人20万英镑至30万英镑的投资，并将允许一系列未来的硬件与乌克里科学和创新的交付有关。商业前的设备将刷新并添加到定期的，可能是年度的基础上。这种敏捷策略旨在利用各个行业的不同方法（例如，一些公司，Nvidia往往具有短（不到3个月）的商业前窗口，而对于其他公司来说，这可能是一年的）。 Excalibur可以使用硬件试验系统来推动整个乌克里研究社区的软件创新。研究人员正确地不愿意花费时间来利用新硬件的优势，这些硬件可能在数年内无法大规模提供，甚至可能证明没有长寿 - 科学领导力要求现在使用研究资金来提供科学成果。此外，狄拉克（Dirac）和其他人将提供资助的RSE努力，以支持开发工作，并在适度的系统中获得新的技术，Excalibur可以降低标准，以参与软件重新设计的过程，并鼓励研究人员进行时间的（适度）投资。在某些情况下，还可能通过利用概念验证系统来进行一些即时科学输出。因此，Excalibur将能够激励整个UKRI研究社区的更大软件创新，并有助于确保当新技术包含在国家服务中时，已经有一些工作流已经能够最佳利用它。这将提高所有乌克里计算服务的生产率，并使英国研究人员能够使用最新的硬件来提供最大，最复杂的计算，从而确保国际领导。