Introduced by Jack Dongarra, they measure how fast a computer solves a dense n by n system of linear equations Ax = b, which is a common task in engineering. The LINPACK Benchmarks are a measure of a system's floating-point computing power. Jack Dongarra, Jim Bunch, Cleve Moler, and Gilbert Stewart
#Goal of linpack benchmark software#
If that delivery time is met, it is conceivable we could see Europe’s first exascale system on the Top 500 in November 2023.For the software library, see LINPACK. These sites will have to provide details for their PUE over the last year for existing systems, depreciation details for the facility, and current electricity pricing, along with the number of staff currently required (and expected) to run a future exascale system.ĮuroHPC JU says any hosting site must meet the baseline requirements in time for the anticipated delivery time of June 2023. Sites will have to have in place power capacity in the 20-25MW range, along with UPS power to cover other system elements (storage, networking) and ample air or liquid cooling capabilities within at least 700 square meters of raised floor space.Ĭenters contending for Europe’s first exascale system will also be ranked by a number of broader factors, including expected TCO of the system in the context of importance of applications, experience of sites in managing large systems, overall physical and IT support infrastructure, and willingness to work with EuroHPC JU users. Only existing, robust sites can apply with such a short deadline as required information includes details from past Top 500 class system experiences. The maximum costs cannot exceed 6% of the overall TCO of the EuroHPC supercomputer. In case the hosting entity decides to include such optional system in their application and its application selected, any grant that will be established to cover the operating costs of each individual EuroHPC supercomputer, may also include a part cover up to 50% of the eligible costs for the development of the advanced experimental platform towards exascale. The procurement of the third EuroHPC pre-exascale supercomputer, MareNostrum5 in Spain, is on-going. Five petascale supercomputers: Vega in Slovenia, MeluXina in Luxembourg, Discoverer in Bulgaria, Karolina in the Czech Republic and Deucalion in Portugal, as well as two EuroHPC pre-exascale supercomputers LUMI in Finland, Leonardo in Italy. To date the EuroHPC JU has already procured seven supercomputers, located across Europe. The development, installation and operating costs of such platform should be marginal to the overall cost of the proposal.” The goal of such a platform shall be to develop an exploratory supercomputing infrastructure for the development, integration, testing, and co-design of a wide range of European technologies suitable to be part of the future European exascale systems. Applications are due by February but EuroHPC has also said that there submissions can include exascale paths versus straight systems, including prototype machines.Īccording to the call, “interested hosting entities may also include in their application an optional system targeting the development of an advanced experimental platform towards exascale systems. The 2022 timeframe to acquire a system (with delivery by mid-2023) is rapidly approaching. In addition, competing sites will need to take those node counts and provide expected sustained Linpack numbers and performance expectations for the domain areas. The straightforward goal, according to EuroHPC, is to build “a capability system with an aggregated performance level capable of executing at least one exaflop of sustained performance measured using Linpack benchmark” and one that covers the needs of a wide range of applications “in particular, grand challenge applications that demonstrably require the capability usage of the supercomputer, i.e., using simultaneously a large part of the resources of the system.”Īnother important target is to be able to “perform at least a Level 1 measurement quality for a Top 500 submission” and centers will be expected to provide their own configuration specs, including types of nodes (with or without accelerators, high memory nodes, etc.) along with the proposed memory, storage, and network architecture along with how those are distributed throughout the system.