The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Authors: Allan Cantle (Nallasway Inc, Open Compute Project), Trent D'Hooge (Lawrence Livermore National Laboratory), Bob Ciotti (NASA), Roger Beeston (ProGrAnalog Corporation), Christopher Chapman (Boyd Coorporation)
Abstract: The Open Compute Project, OCP, is developing a set of interoperable HPC building blocks addressing the convergence of big data, artificial intelligence and HPC.
The aim of this BoF is to introduce OCP HPC Subproject HPCM, High Performance Computing Module, to the supercomputing community as an efficient and cost effective option for many workloads. The BoF panelists will review the OCP HPC Sub-project roadmap, giving a holistic perspective on Heterogeneous design, Data locality, Interconnect flexibility, efficiency, Power delivery, and Thermal Management. A Q&A discussion with audience participation will follow, covering how to successfully bring this HPCM concept to reality.
Long Description: CHALLENGE
HPC Class applications continue to broaden beyond traditional HPC workloads with the convergence of big data, artificial intelligence and high performance computing. There are also seismic changes in the building blocks of computing architecture from a multitude of new processors and accelerators, to the blurring of the boundaries between memory and storage media.
There is no “one size fits all'' HPC solution for every application domain, but instead each application needs a particular HPC Domain Specific Architecture, DSA, to achieve the highest performance at the best efficiency and lowest cost.
Historically, the Top 50 systems of the Top500 list have been expensive and used proprietary architectures while many of the other systems use commodity off the shelf technology. With many innovations mentioned previously the HPC industry would benefit from a wider set of commodity building blocks.
The OCP HPC Sub Project has recognized that the community desperately needs industry standard solutions that emphasize the features of modularity, flexibility and composability to deliver an excellent performance to price ratio.
SC21 PANEL - HARDWARE CO-DESIGN ACROSS TRADITIONAL SILO'S
The OCP HPC Sub Project has an opportunity to restate the physical building blocks of computing to dramatically improve performance and simultaneously reduce power and cost. Concepts being considered have taken a clean slate approach and analyzed the modularity, flexibility and composability needs from a holistic, systems engineering perspective. To be successful, a hardware co-design approach is required between experts in the following areas : systems, silicon, packaging, signal integrity, power, thermal and mechanical. This SC21 BoF will include a panel of experts from these disciplines who need to resolve their specific technical challenges while appreciating the challenges of the other parts of the system and working together towards an optimal solution.
The ideal composable building blocks would allow a systems architect to quickly design and build supercomputers with the right balance of resources for particular applications. Ideally, one will be able to easily construct architectures with a range of metric ratios, such as floating point operations to memory bandwidth to network latency to storage capacity.
An overall measure of success would show that a range of applications with different resource needs could all be specifically accommodated with best performance, efficiency and cost for a range of build scenarios.
The BoF Session will discuss how the OCP HPC SubProject has successfully worked across traditional Silo's to create fully open architectural building blocks of all computing components. The effort will be shown to be a work in progress effort and will identify and discuss challenges that need to be overcome for the HPC Module, HPCM, to become the industry standard way of building Exascale HPC Machines.
URL: https://www.opencompute.org/wiki/HPC
Back to Birds of a Feather Archive Listing