The Competitive Edge of POWER9 for Cloud and AI
Terry Thomas, director of Cognitive Systems Enterprise Offerings, IBM, shares how POWER9 is purpose built for cloud and AI.
Image by Jason Griego
By Jim Utsler03/01/2019
Using a system built by a single vendor from the ground up has its benefits. For example, both the hardware and software have been optimized to essentially function as a single entity, removing the complexity involved in multivendor scenarios.
Such is the case with IBM POWER9* servers, which make a compelling argument for upgrading from POWER7* or POWER8* processor-based servers. Tightly integrated improvements to CPUs, GPUs and I/O—as well as security—create an environment in which users of (or those interested in) emerging technologies such as artificial intelligence (AI) and on-premises cloud solutions can thrive.
“We intentionally delivered numerous innovations within the entire POWER9 stack—software and hardware—with a focus on meeting the demands of cloud computing and using AI to derive faster insights. So, clients are now using POWER9 to build or modernize their multicloud environments while also gaining unprecedented insights from their data,” says Terry Thomas, director of Cognitive Systems Enterprise Offerings, IBM.
The Promise of Insight
The POWER9 processor is the foundation of the No. 1 and No. 2 supercomputers in the world—Summit and Sierra, respectively—a testament to how cutting-edge technology should be employed in the computing environments of today and tomorrow. For example, the chip includes NVIDIA NVLink, OpenCAPI and PCIe Gen4. Built with data-intensive, AI and machine learning workloads in mind, POWER9 servers provide stark performance advantages when compared to Intel*-based servers, with POWER9 cores having 2x more threads than x86 cores.
POWER9 servers also cut time-to-insight with up to 16 TB (E950) or up to 64 TB (E980) of buffered DDR4 memory for in-memory databases; they speed up applications with 1.8x the memory bandwidth compared to x86 servers; and they respond more rapidly and with greater agility to changing business requirements and growth with Capacity on Demand for processors and memory (ibm.co/2OQqKka).
Which may, in part, be why information-rich and hungry organizations have turned to POWER9. For instance, Uber plans to use the Summit supercomputer at Oak Ridge National Laboratory to test an AI project at scale, across 202,752 POWER9 cores and 27,648 NVIDIA GPUs.
Although AI goals differ, they all desire a common attribute when it comes to machine and deep learning: speed.
A Complete Set of Tools
“Within PowerAI, we optimized open-source software like TensorFlow, Caffe and Torch to take advantage of our enterprise AI servers and high-speed NVLink CPU/GPU connections,” Thomas remarks. “Our processors, NVIDIA GPUs and optimized AI software, such as IBM Snap Machine Learning, enable data scientists to train large data workloads 46x faster than they would if they were using Intel-based servers.”
PowerAI also comes with distributed deep learning libraries that enable TensorFlow to run hundreds of CPUs across tens of servers, helping reduce AI-model training time from weeks to hours. And then there’s the point-and-click PowerAI Vision, which allows laypeople to quickly prototype and build high-quality image and video models for a variety of AI applications.
PowerAI Vision automatically trains AI models to leverage vision technologies for image and video detection, allowing users to reference and label specific images so training models can go through existing data pools and quickly and automatically identify and label other data within the same pools.
“Users who upgrade to the new POWER9 portfolio of systems can seamlessly plug into cloud environments with simplified multicloud capabilities."—Terry Thomas, director of Cognitive Systems Enterprise Offerings, IBM
“Years ago, clients’ system value metrics were largely based on CPU capabilities: How many transactions can I drive? We call this the ‘process economy.’ Now, clients’ system value metrics are shifting to insights: How fast and what types of insight can I gather from my data? We call this the ‘insight economy,’ ” Thomas says. “Recognizing clients’ needs, along with changes in the IT landscape—the effective end of Moore’s law, new AI approaches and algorithms, innovations in semiconductors, the use of GPUs and high-speed interconnects—with POWER9, we’ve basically disrupted typical approaches to computing by reinventing compute to realize the promise of insight.”
Economies of Scale
This type of disruption applies to cloud, as well. PowerAI is, in fact, very comfortable running there, whether on- or off-premises, in hybrid or multicloud, or co-located environments. Scaling with machine learning, deep learning or distributed-deep learning fits seamlessly within cloud environments.
This has been made possible thanks to not only IBM’s strategic partnership with VMware, with its vRealize Suite of multicloud management and application monitoring solutions, but also because of the OpenStack-based IBM Cloud* PowerVC Manager. This tool allows users to provision workloads and manage virtual images across cloud environments. Cloud PowerVC Manager also provides a private cloud portal for self-service provisioning, and the PowerVC API provides an open integration layer for ecosystem partners such as SAP and VMware.
“Users who upgrade to the new POWER9 portfolio of systems can seamlessly plug into cloud environments with simplified multicloud capabilities. They can dynamically scale, compute and use memory on demand to build cloud designs for their most data-intensive workloads, with open integration with leading multicloud managers,” Thomas notes.
AI and cloud are only part of the POWER9 story. For example, a user running a four-node POWER7 770 with 64 cores and 1 TB of memory can consolidate to a one-node POWER9 E950 with 24 cores and 1 TB of memory. The gain is nearly 3x more per core performance and a 60 percent reduction in cores, lowering software-licensing fees and freeing up money for future development.
Data center floor space and system management are also both reduced, lowering energy usage and allowing for the reallocation of personnel to work on other projects. Additionally, IBM’s Capacity on Demand capabilities simplify consumption-based infrastructures, with users paying only for the resources they require with almost instant response to new workload demands.
The built-in IBM PowerVM, the Power Systems hypervisor for data and mission-critical applications, also allows users to dynamically scale compute and memory on demand. This can be based on business needs, with POWER9 providing accelerated and encrypted live VM mobility.
“Our Power E950 and E980 systems offer new economy of scale and flexibility to enable clients to better manage their business within the complex computing environment of today’s large enterprise,” says George Gaylord, offering manager for high-end Power* servers, IBM. “POWER9 systems can also integrate seamlessly into heterogenous environments managed with VMware’s vRealize suite.”
Security is also a critical part of the Power Systems architecture. It has built-in security at the chip level that spans the entire stack, including in the processor, the OS and everything else in between. In fact, these systems are designed to deliver end-to-end security features, beginning at processor boot time.
As Thomas explains, “One technology I’d like to specifically call out is the encryption engines within our processor. We doubled those on POWER9. That means customers, whether their workloads are mobile or static, are assured of high security and integrity of their workloads. This is specifically significant for clients that leverage LPM—live partition mobility—as they move workloads across their environments.”
Powerful and Flexible
Thanks to recent and pertinent discussions about AI and cloud, upgrading from POWER7 or POWER8 to POWER9 seems like an easy choice. Thanks to denser CPUs, integrated and accelerated GPUs (the POWER9 processor fully exploits the capabilities of its GPU accelerators at 4x the bandwidth of POWER8), and increased I/O and security, POWER9 may indeed represent the sweet spot for many Power Systems users, no matter the size of their organizations.
In fact, POWER9 servers come in a variety of shapes and sizes. For example, the AC922, an accelerated compute server, delivers unparalleled performance for modern high-performance computing, analytics and AI. The L922 also handles AI and cloud, but exclusively runs Linux*. The IBM Power System H922 is more closely aligned with shops running SAP HANA. And the E980 and E950 provide a unique blend of enterprise-class capabilities for large and medium enterprises.
Whatever the POWER9 configurations users might choose, they can rest assured that they won’t be missing out on the latest computing trends, including AI and cloud. Additionally, these systems excel at running critical day-to-day workloads. This is, in large part, thanks to the hardware/software stack IBM has baked in and optimized for every POWER9 server.
“We intentionally structured our POWER9 offerings and capabilities so they can seamlessly plug into most environments without the need of a unique software stack or any other capabilities built on top of POWER9,” Thomas notes. “That’s what makes the platform so powerful and flexible.”
AIX / IBM i / Linux on IBM Z / Article / Cloud / Systems management / Artificial intelligence / Executive perspective / IT infrastructure / Machine learning / Power Systems / POWER9 / Cloud strategy / Hybrid cloud / Private cloud / Data security / Workload management / Multicloud
Jim Utsler, IBM Systems magazine senior writer, has been writing for IBM since the mid-1990s.