IBM’s POWER9 Targets AI, ML and Cognitive Workloads

April 20, 2018

IBM’s POWER9 Targets AI, ML and Cognitive Workloads

By Jean S. Bozman

IBM’s POWER9 processor has emerged as an engine for Artificial Intelligence (AI), Machine Learning (ML) and Cognitive Computing. Long anticipated on the IBM processor roadmap, POWER9 was announced in Q417, and is already moving into production at IBM Power Systems customer sites.

Through the OpenPOWER consortium, partners will leverage IBM’s POWER9 technology, which is moving into new types of uses, notably at Google, in China and at HPC data centers (Note: Please refer to the accompanying OpenPOWER blog, which examines these market segments).

Historically, POWER processors were used for enterprise workloads, like SAP/ERP and IBM DB2. Customers looked to new and faster processors, every few years, to boost performance for computation and storage. They continue to support those workloads, but that appeals most to IBM’s longtime installed base.

The newest pull on IT managers is the “presence” of these new POWER9-based systems. Now, IBM’s emphasis on AI, ML and Cognitive workloads is inviting new participation in the POWER/OpenPOWER architecture, and its support of new and emerging applications will attract a wider customer base worldwide. In contrast, a more traditional business model might have kept POWER processors within the IBM customer installed base.

This approach 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, as the foundation for new business models and new customers.

 

The Workload Mix

What will the workload mix for POWER9 and OpenPOWER look like? Examples include IBM Watson-enabled cognitive applications; high-performance computing (HPC) for deep scientific analysis; and data-intensive enterprise workloads like SAP supply-chain logistics and blockchain opportunities (e.g., IBM-Maersk joint venture workloads for international shipping).

IBM plans to build differentiated AI/ML solutions to market, across many industries (e.g., financial, retail, oil/gas), starting this year. To signal this change in the marketplace, IBM’s POWER-enabled server systems, marketed as IBM Power Systems, are now included in the IBM Cognitive Systems business unit. This approach emphasizes the IBM POWER9-based systems’ role for cognitive and AI workloads.

The key element of these systems is that POWER9 processors will run data-intensive applications and business analytics that benefit from optimized CPU-to-GPU throughput resulting from Nvidia’s GPUs embedded on the POWER9 processors. Data-intensive deployments on POWER9 will also leverage high-speed I/O interconnects and Nvidia’s NVLink II technology, which is built into, or embedded in the POWER9 processor itself.

 

POWERAI

For the intensive analytics applications, POWERAI is a software framework for running AI/ML and Cognitive workloads running on POWER9 and OpenPOWER.

POWERAI for analytics is enterprise software that is targeted at deep learning problems. Customers use POWERAI to exercise the CPU-to-GPU NVLink II interconnects. When used in parallel, with multiple processors, POWER AI allows businesses to load larger deep-learning models – and to “train” them to sort the data, than was possible before.

 

SAP and Enterprise Applications

Enterprise workloads for the POWER9 are clear, as well. These include SAP HANA workloads supporting in-memory processing and large relational databases like IBM DB2. Both SAP HANA for in-memory analytics and the IBM DB2 relational database analyze large amount of data – and both workloads benefit from high-speed data handling.

 

The Strategy

IBM strategy is to have a differentiated hardware strategy centered on selected workloads that benefit from embedded GPUs on POWER9 processors. A newly announced POWER9-based server, IBM ACC922, is being marketed for data analytics, AI and ML workloads – and for HPC workloads that traditionally ran on expensive scalable supercomputers.

POWER and OpenPOWER processors represent just a small slice of the overall processor marketplace, compared with the high volume of shipments for x86 systems from Intel. We should note that x86 systems also leverage multiple Nvidia GPUs, but the implementations are different than in the systems based on POWER OpenPOWER processors. So far, the Nvidia GPUs are not embedded in x86 CPUs – but rather leverage NVLink connections between processors and CPUs.

 

OpenPOWER Ecosystem to Grow the POWER9 Technology Base

We expect that ripple effects should follow from POWER9 technology, because IBM is expanding the OpenPOWER ecosystem by involving dozens of systems vendors, networking companies and software providers. Linux support, including the ability to run little-endian Linux applications that were written for x86 platforms, will broaden the POWER9 use cases.

Deliberately using an open licensing strategy with the OpenPOWER consortium, IBM clearly understands that it needs partners and a wider customer base to keep the POWER designs moving forward. Otherwise, a shrinking user base for traditional RISC chips would have limited POWER’s opportunities in the marketplace, and brought the roadmap to a close.

This pattern of cooperative design with partners combines IBM’s CPUs with Nvidia’s GPUs – and includes Google, and Asian system vendors in bringing OpenPOWER partners (LINK to OpenPOWER partners) to a broader marketplace than IBM would have accomplished with a single-vendor strategy. Importantly, OpenPOWER members have a role in developing new specifications for POWER processors. Another avenue for cooperative design is open-source software, such as OpenCAPI, with open APIs that access coherent memory associated with the POWER processor designs.

As discussed at the OpenPOWER Summit in March, 2018, POWER9 will be leveraged in a broad mix of use-cases, extending beyond IBM’s customer base. Having a broader base and more use-cases provides a better business model – and is already justifying a longer roadmap for POWER.

As the OpenPOWER Summit made clear, IBM is already working on POWER10, which we expect will be announced by 2020, and deployed in later years. By way of background, POWER7 shipped in 2010, followed by POWER8 in 2014 and POWER9 was introduced in 2017.

 

Cloud Computing, Jean Bozman , , , , , , , , , ,
About Jean Bozman

Jean is a senior industry analyst focusing her research on server technology, storage technology, database software and the emerging market for Software Defined Infrastructure (SDI).