IBM is aggressively targeting the AI enterprise market with dedicated AI hardware featuring fast inference performance and low-latency responsiveness.

IBM announced on the 28th that it has officially launched 'Spyre Accelerator,' a high-performance AI accelerator for scaling generative and agent-based AI tasks.

The product is compatible with IBM's mainframe systems, IBM Z17 and IBM LinuxONE 5, and will also be available for Power11 servers starting in early December.

Spire Accelerator provides fast inference performance and low-latency responsiveness, and is designed with the security and resilience of core data and trading systems as top priorities. IBM emphasized the need for dedicated hardware capable of securely running AI in on-premise environments to address the transition from existing logic-based workflows to AI agent-centric real-time inference environments.

This product is the result of joint development between IBM Research’s AI Hardware Center and the Infrastructure Business Unit. Its completeness was enhanced through a rapid iterative development process in collaboration with the Yorktown Heights Research Center and the University of Albany’s Center for Emerging AI Systems. The commercial product is provided as a System-on-Chip (SoC) featuring 32 individual acceleration cores and 25.6 billion transistors, offered as a 75W PCIe card based on a 5-nanometer process. Up to 48 clusters can be configured on IBM Z and LinuxONE systems, and up to 16 on Power systems.

Barry Baker, COO of IBM’s Infrastructure Business Unit, stated, “Through the Spire Accelerator, IBM systems have gained the capability to support multi-model AI,” adding, “It helps scale AI-driven core business operations while maintaining security, resilience, and efficiency.”

“We are proud that the first chip from the IBM Research AI Hardware Center, which has been responding to the growing demand for AI computing since 2019, has entered the commercialization phase,” said Mukesh Kare, Vice President of IBM’s Semiconductor and Hybrid Cloud division.

IBM customers can leverage AI acceleration capabilities based on on-premises systems through the Spire Accelerator, and in particular, on IBM Z and LinuxONE systems, it can be used with the Telum II processor to apply multiple AI models to predictive tasks such as advanced fraud detection and distribution automation. Power-based servers accelerate data transformation and document integration for generative AI through the AI service catalog, providing performance capable of integrating over 8 million documents per hour into a knowledge base.

More information about the Spire Accelerator can be found on the official IBM website.