By Jean S. Bozman
IBM is expanding the base of customers who will deploy the IBM POWER9 processor and OpenPOWER technologies, by extending its POWER architecture for use in new and emerging workloads.
This approach, born with OpenPOWER’s inception in 2013, is building a larger ecosystem around POWER than would have been possible if IBM had held onto POWER processors as products for IBM customers only. It’s worth noting that OpenPOWER members have an open hardware license that allows them to innovate on top of the IBM POWER9 design, and to fabricate the resulting processors outside of IBM’s own POWER9 manufacturing process. [NOTE: Please refer to the accompanying blog regarding POWER and OpenPOWER].
In March, 2018, as the OpenPOWER Summit was held simultaneously with the IBM THINK conference in Las Vegas, it was clear that OpenPOWER technologies would be giving rise to a wide variety of born-on-the-cloud and emerging applications in data-intensive analytics, AI, ML and Cognitive workloads.
Following are three approaches underway to implement OpenPOWER in different market segments – Google, China and HPC (high performance computing):
Google’s Strategy for OpenPOWER
Google’s interest in OpenPOWER and POWER9 processors is coming into focus. Google has a special role in the OpenPOWER consortium — a Google executive was chair of the OpenPOWER board of directors, with continuing Google board membership. Google’s presentation at the OpenPOWER Summit explained that OpenPOWER-based infrastructure will support selected workloads in Google’s data centers.
Rather than planning to displace the x86 server architecture Google uses to run its key workloads—Google Search, Gmail and Google Cloud—Google plans to apply POWER9 architecture for special workloads that will benefit from POWER9-based clusters and high-speed CPU-to-GPU NVLink interconnects combining OpenPOWER and Nvidia technology.
Google has already built, and put into production, a large cluster of Google “Zaius” servers, which are based on OpenPOWER processors and Google-made motherboards. The server nodes support high-capacity memory; the OpenCAPI interface (SPIs) for networking software; PCIe Gen4 I/O interconnects; OpenBMC management software – and housed in OCP-compatible Open Rack chassis.
While Google stopped short of providing details about the Zaius server clusters – in effect, a scalable Linux supercomputer – the net result will be a system for AI/ML and Deep Learning powered by Google’s own TensorFlow software. We believe the POWER9-based Google systems should be viewed as a resource for the Google Cloud Platform (GCP) – and as a platform for Google’s internal use in R&D.
OpenPOWER and China
Opportunities in the China market are an important factor in expanding the OpenPOWER ecosystem among Chinese systems vendor firms. Workloads of high interest among these Chinese companies include hyperscale computing for Cloud Service Providers (CSPs); HPC for scientific purposes; and deep analytics for financial services and banking institutions throughout China.
A number of Chinese vendors, including Inspur, Digital China, Wistron and others, are planning to use OpenPOWER processors in their systems for data-intensive analytics and high-performance computing (HPC) uses.
The open hardware design process of OpenPOWER allows vendors outside IBM to manufacture their own OpenPOWER designs. This strategy opens an avenue for innovation on top of IBM’s POWER9 architecture and chip design. The optimization process provides a platform for customized innovation, resulting in a range of POWER9 implementations. Outside China, other Asian companies are also involved in OpenPOWER, including Hitachi and NEC in Japan, and SK Hynix, Samsung in Korea, among others.
HPC (High Performance Computing)
Universities and scientific research labs are listed as partners in the OpenPOWER consortium, part of the high performance computing (HPC) customer base for POWER and OpenPOWER. In the U.S., two of the largest national research labs, Argonne and Oak Ridge are building the Summit supercomputer, with thousands of POWER9-based server nodes connected in clusters.
Universities that are participating in OpenPOWER include CalTech in southern California, LSU in Louisiana, Oregon State University, the University of Michigan and Rice University in Texas. Universities in Spain, the Netherlands and Germany are also contributors to OpenPOWER.
High performance computing (HPC) relies on fast processors, large cache memory and main memory, and high-speed interconnects to process large datasets quickly. The CPUs, GPUs and high-speed links from Mellanox and others are all components of cluster-centric HPC supercomputing systems.
IBM is proving that POWER9 is a high-value chip, even if it is not a high-volume chip. IBM is bringing POWER9 to market, creating a portfolio of POWER9-based products — and continuing its POWER roadmap past the Year 2020.
Importantly, IBM is expanding the OpenPOWER ecosystem of system vendors and customers leveraging OpenPOWER for AI/ML and cognitive workloads, creating a broader base for what once was a highly niched RISC processor that began its technical timeline some 20 years ago. In recent years, the milestones for POWER were: POWER7 processors shipped in 2010, followed by POWER8 processors in 2014 and POWER9 processors in 2017.
This sheds light on IBM’s view that technology developed in its IBM Research labs and R&D operations can be leveraged in the wider IT world to support on-premises and off-premises workloads. IBM’s strategy for POWER and OpenPOWER depends on IBM’s ability to bring differentiation in workload acceleration and AI capabilities, to a broader group of IT industry vendors. IBM must keep growing the OpenPOWER ecosystem among cloud services providers, enterprise customers – and more partners in the IT industry.
POWER, which began life as a RISC chip for Unix systems, is accomplishing much more than it would have as a single-vendor follow-on product for RISC-based computing with AIX and Linux. As a platform for AI and analytics enablement, that IBM’s POWER initiative is much more than that, as an engine for future systems.