Intel has used this year’s Supercomputing Conference to officially launch its Omni-Path high-speed interconnect fabric and disclose details of an Xeon Phi workstation it plans to make available for developers. However, the Knights Landing processor is now not set for general availability until next year.
The Supercomputing Conference (SC15), or the International Conference for High Performance Computing, Networking, Storage and Analysis, to give the event its full name, is a traditional venue for Intel to detail updates to its high-performance computing (HPC) portfolio.
This time, Intel was keen to stress that it has shifted focus to a holistic system-wide strategy in order to break through the bottlenecks that hold back the development of ever more powerful HPC systems.
“One of the reasons we are transcending from being a pure compute or processor provider to taking a more system-wide view is that there are a number of well-known bottlenecks in today’s supercomputing systems that are only going to get worse over time, so to keep moving you really have to take a more systemic view and look at memory, I/O, storage and even the software stack,” said Charles Wuischpard, vice president of the Data Centre Group and general manager of the HPC Platform Group at Intel.
Customers are also increasingly asking for a single architecture that can support big data, visualisation, machine learning and HPC workloads instead of requiring dedicated infrastructure for each, Wuischpard added. This led to the creation earlier this year of Intel’s Scalable System Framework (SSF) that can serve as a flexible architecture for HPC systems scaling from a single rack up to the largest supercomputer.
A key part of SSF is the Omni-Path fabric technology, officially launched at SC15, which provides the connection between individual nodes in an HPC cluster. This has been developed to meet the divergent requirements of delivering high performance while keeping overall cost down, according to Intel.
“The industry has been waiting for this anxiously for some time,” said Wuischpard. “This is a 100Gbps solution that’s really tuned for true application performance, and we’ve measured up to 17 percent lower latency and a 16 percent higher messaging rate than InfiniBand EDR.”
Intel is making available not just Omni-Path Host Fabric Interface adapters, but a 48-port switch chip and switches based on it which enable network topologies at lower cost using fewer components, according to Wuischpard.
“If you’ve got a budget of, say, $1m you should be able to spend 26 percent more of that now on compute or other parts of the architecture, rather than on the fabric,” he claimed. Omni-Path also includes features to provide greater resiliency without degrading overall system performance, he added.
Omni-Path has been designed so it can be implemented using copper or optical fibre cabling, and is also set to be integrated into Xeon Phi processor chips at some point, starting with versions of the Knights Landing chip that was originally slated to be available before the end of 2015.
However, Intel revealed that it does not yet have full production-level general availability of Knights Landing, although there are systems installed at some customer sites based on the chip already. These include the Sandia National Laboratories in the US, the French Atomic Energy Commission and the Cray supercomputer company. Full general availability will now come in 2016.
Knights Landing is designed to feature up to 72 cores, each based on an architecture similar to Intel’s Atom processors, and is due to be followed by a successor called Knights Hill that will be manufactured using a 10nm production process.
Another surprise disclosure from Intel is that the firm plans to deliver an Xeon Phi-based workstation that will enable developers to have their own local system for building and testing applications before deploying them to the data centre.
This is based on Knights Landing, the first in Intel’s Xeon Phi family designed to operate as a standalone CPU rather than a co-processor plugged into a PCI Express slot in an Xeon-based system.
“The plan is to ship a workstation that I kind of think of as the ‘ninja’ workstation, which will be Xeon Phi with the full software suite pre-tuned and pre-loaded, along with code samples and all the tools you might need as a developer to run, test and benchmark your own code, and this will be rolling out in the early part of 2016,” Wuischpard said.
Intel also announced new versions of its software-defined storage solutions based on the high-performance open source Lustre file system. Intel Enterprise Edition for Lustre v2.4, Intel Cloud Edition for Lustre v1.2 and Intel Foundation Edition for Lustre v2.7.1 are all slated for availability before the end of this year.
Wuischpard said that, while Lustre on its own does not generate much revenue for Intel, it is “absolutely strategic” in enabling full system performance in its supercomputer platforms.
“Again, this is part of us taking a full end-to-end view, and Lustre is really starting to move from more of a lab/power user mode into more of an enterprise mode now,” he said.
Finally, Intel said it will offer supported versions of the OpenHPC Software Stack to customers from 2016 as part of the SSF.
Managed by the Linux Foundation, OpenHPC was announced last week and is “an open source, stable and collaborative software stack for HPC that everyone can use”, according to Wuischpard.
“At the end of the day, everyone just wants a working software stack, so there is broad support from the industry, from ISVs, system builders, integrators and even end users to have one common software stack to work from,” he explained.