Spyre AI Accelerator Now Available on Power11

The card became generally available Dec. 12, expanding on-premises inferencing abilities for organizations racing to operationalize AI

Six weeks after IBM’s Spyre AI accelerator began shipping for the Z platform, the PCIE-attached card is now generally available on Power11 as well.

Complementing Power11’s own on-chip AI accelerator, each Spyre chip comes with 32 AI accelerator cores, 128 GB of LPDDR5 memory and compute exceeding 300 TOPS, according to IBM data. A PCIe Gen4 I/O Expansion Drawer can accommodate up to eight Spyre cards, amounting to 1 TB of memory.

A Full-Stack AI Strategy

IBM executives showed off the new hardware during a press briefing at their Poughkeepsie, New York, campus in early December, outlining how the accelerator fits into an approach to AI that emphasizes full-stack integration over pure processing power.

“We’re not competing on just raw horsepower,” said Bargav Balakrishan, VP of product management for IBM Power. The Nvidias of the world have them beat on that front, he admitted. “But what we can do better is that integration move,” he said, citing the “turnkey simplicity” of the Spyre card.

This integrated, plug-and-play approach to AI helps organizations address the vexing challenge of innovating while keeping the business running smoothly, an age-old problem for every customer IBM speaks to, Balakrishan said.

AI ROI Remains Elusive

Organizations are struggling with large-scale AI implementations—one oft-cited MIT study found that 95% of organizations are getting no return on their AI investments. So while convenience and expedience remains a hallmark of AI for the end user, implementing it at scale is a different matter.

To operationalize AI at scale, organizations must consider factors like network configuration, business workflow integration, data management and potential skills gaps, Sebastian Lehrig, worldwide AI on IBM Power leader, noted in his recent appearance on Charlie Guarino’s TechTalk SMB podcast. Large-scale AI implementation, he continued, requires an interdisciplinary approach involving expertise in data science, software development, security and perhaps MLOps.

“With Spyre, we reduce that complexity to the bare minimum and essentially launch, now, a whole catalog of pre-built AI services that you can install with a single click on Power,” Lehrig said.

Protecting Data

Another challenge in successful AI implementation has to do with data security and sovereignty, with organizations hesitant to send sensitive data like personally identifiable information to the cloud for processing. By bringing more AI oomph to Power11 servers, Spyre helps organizations keep all that data on-premises. And with the provided industry-standard APIs, AI applications can be tested in the cloud before being moved back on-premises for production, Lehrig explained.  “As soon as you move to production, often it’s easier to also move the AI then to where the data lives,” he said.

Broadly, Spyre’s two main uses are transaction monitoring and productivity enhancement. Complementing on-chip AI accelerators and the predictive models they are best suited for, Spyre puts generative AI capabilities near the core and allows for multiple model types to be used for fraud detection on a single transaction.

On the productivity side, IBM is pointing to Spyre’s ability to help power agentic AI and AI assistance in products such as Project Bob, the AI-enabled integrated development environment (IDE) that is expected to be available in Q1 of 2026.

Agentic frameworks are not limited to their IT environment, however; they can also extend into the real world. One convenient example can be found in the trucking industry, where agentic AI is already in operation, Balakrishan said.

In an agentic AI system tuned for shipping, when a truck has a problem on the road, AI agents can pull relevant data like the truck’s route and destination, find the nearest repair station, schedule an appointment and even check to make sure the shop has the necessary parts on hand, Balakrishan said.

With Spyre now generally available for Power11, those trucks have some precious cargo to carry.

Preparing for Spyre

To install a Spyre card in a Power system, the following must be present:

Hardware

• IBM Power11 processor-based server
• A minimum of 1 TB of memory

Software

• Red Hat Enterprise Linux 9 for Power LE version 9.6, or later
• Spyre Enablement Stack
• IBM Spyre for Power Entitlement
• Red Hat AI Inference Server

---