The hallmark of the cognitive enterprise will be optimized technology stacks that meet the demands of AI workloads – and that must start with hardware.
As the industrial age gives way to the digital era, automation and various forms of artificial intelligence (AI) are driving a fundamental shift in how enterprises engage with customers, deliver services and optimize business functions.
Leading organizations are using AI to enable rapid decision making, capture and codify value-driving organizational expertise, and to deliver actionable insights at the point of engagement. The application of these technologies demands that organizations reshape both their organizational models and their technology stack to transform themselves into what we call a Cognitive Enterprise.
While most enterprises are at some stage of their technology transformation, many of them are taking too narrow a view and looking only at the software layer as they do so. This limited focus may prove to be a mistake, as the demands of the cognitive enterprise require a reimagining of the stack from hardware on up.
The industry is in the throes of a movement toward ‘software-defined’ architectures. The rationale is compelling: abstract the software that provides logic, orchestration, and management from the physical infrastructure and you get vastly improved efficiency and agility.
As a result, this software-defined approach was one of the chief enablers as cloud companies deployed what we now call web-scale architectures.
Understandably, enterprise organizations have been frantically trying to replicate these approaches as they strive to transform themselves. There are, therefore, now viewing infrastructure as a pure commodity and focusing all of their efforts on software optimization.
Because of their compute-intensive nature, however, AI workloads require an optimized stack that extends beyond the software layer and which optimizes both the underlying hardware as well as the integration among software components.
Web-scale companies have already realized this dependence on optimized hardware and have therefore been developing proprietary, purpose-built, and fully integrated architectures from the hardware layer on up to power their demanding AI workloads.
Enterprise organizations are likewise finding that their AI workloads are driving exponential growth in both data and compute demand and they will, therefore, need to follow suit.
As organizations begin to develop their AI-optimized stack, they quickly realize that commodity hardware is insufficient for the optimization they seek. Such optimization must start at the processor level with a laser-focus on removing bottlenecks, increasing core performance, expanding memory bandwidth and enabling compute acceleration.
But creating an AI-optimized stack is about more than just having optimized hardware. They must also tune their entire software stack to meet the particular needs of AI workloads.
This software optimization must occur at two levels. First, the software must be optimized to work with the specialized hardware to fully take advantage of these purpose-built hardware innovations. Second, the software must be configured to work together as efficiently as possible in the context of workload demands.
As enterprise leaders transform their organizations into cognitive enterprises, they will need to differentiate workloads based on both business value and the demands they place on the technology stack.
The most critical to their transformative efforts will be these intensive AI workloads, which will leave them seeking the most effective ways to rapidly build and deploy stacks optimized to power them.
IBM has developed fully integrated and vertically-oriented solutions to help enterprise organizations more rapidly deploy these AI-optimized architectures and realize value more quickly. Its approach includes purpose-built hardware solutions based on its new POWER9 chip, along with a full complement of open source packages IBM has optimized for its hardware and to meet the demands of AI (what it calls its PowerAI offering).
Whether or not organizations use these sorts of pre-built solutions, their objective must be to focus on optimizing the stack for the particular demands of AI workloads and then ensure that they can rapidly deploy these architectures to meet dynamically changing business demands. The ability to do so will be one of the hallmarks of the cognitive enterprise.
Copyright © Intellyx LLC. IBM is an Intellyx client. Intellyx retains full editorial control over the content of this paper.