IBM today at the Hot Chips 2024 conference revealed that next year it will be adding a 5.5 GHz IBM Telum II Processor with eight cores that is better optimized to run artificial intelligence (AI) inference engines on IBM X mainframes.
In addition, IBM is making available a technical preview of an add-on; a previously announced IBM Spyre Accelerator chip that makes it simpler to invoke multiple AI models simultaneously in 2025. Attached to a mainframe using a PCIe adapter, the IBM Spyre Accelerator provides access to 1TB memory across 32 core enables mainframes to run multiple machine learning or deep learning AI models alongside large language models (LLMs).
These additions to the venerable IBM platform will make it simpler for applications to invoke the most appropriate type of AI model for their use case using a platform that also has an IBM Telum II Processor that provides access to 40% more caching, 360M of memory and a revamped data processing unit (DPU) to provide 50% higher I/O throughput, says Christopher Berry, distinguished engineer for microprocessor design at IBM.
The overall goal continues to be to make it easier to run AI models on either z/OS or Linux ONE operating systems where most of the mission-critical and analytics data an organization relies on is already located, he adds. Otherwise, organizations will have to devote additional resources to moving that data to another platform rather than taking advantage of the processors IBM is adding to the mainframe that are specifically optimized for AI models, notes Berry.
Manufactured for IBM by Samsung Foundry, these processors also provide a more energy-efficient approach to run large AI models, says Berry.
In general, the processors are part of a larger IBM effort to streamline the design of mainframes in a way that makes them more accessible to a wider range of IT professionals. “We’re focused on system simplification,” he says.
The cost of entry for a mainframe remains considerably higher than traditional servers, but the total cost of ownership (TCO) for a mainframe remains appealing for larger enterprises, especially if they are consolidating Linux and z/OS workloads using a pricing model that is based on the total amount infrastructure consumed at peak levels.
The biggest challenge, however, continues to be finding the expertise required to manage mainframes as many of the IT professionals with these skills continue to retire. IBM, as a result, is making a concerted effort to make the mainframe accessible to IT operations teams of all skills levels, while at the same time increasing the number of Java and Python applications running on the platform.
In the meantime, it’s still early days so far as the deployment of AI inference engines on IT platforms is concerned, but IBM is making a case for running multiple types of AI models on mainframes. Whether that leads to more mainframes being employed remains to be seen, but as the forces of data gravity continue to exert themselves, the one thing that is for sure is a lot more AI models will be running on a platform that is now more than a half-century old.
