IBM’s plans for its Power server family promise faster and more programmable AI capabilities along with a new, more powerful processor. Highlights of the 2025 roadmap include:
- IBM’s Spyre Accelerator for AI will be part of the Power processing systems.
- The Power server line will be anchored by a new processor, the IBM Power11.
- A new AI-based assistant will aid in RPG application modernization and development.
“Jointly designed by IBM Research and IBM Infrastructure, Spyre’s architecture is designed for more efficient AI computation. Notably, the chip will send data directly from one compute engine to the next, leading to an efficient use of energy,” wrote Bargav Balakrishnan, vice president, product management, IBM Power, in a blog about the roadmap.
“This family of processors also uses a range of lower precision numeric formats (such as [AI datatypes] int4 and int8), to make running an AI model more energy efficient and far less memory intensive,” Balakrishnan stated.
According to IBM, the Spyre Accelerator significantly boosts speed and accuracy of AI processing capabilities for the Power server. By offloading AI tasks to the accelerator, servers can achieve performance gains, leading to faster response times and improved overall system efficiency.
IBM earlier this year introduced the Spyre Accelerator, saying it would be included with the forthcoming Telum II processor set to debut in its Z and LinuxONE mainframes next year. The Spyre Accelerator will contain 1TB of memory and 32 AI accelerator cores that will share a similar architecture to the AI accelerator integrated into the Telum II chip, according to IBM. Each Spyre is mounted on a PCIe card.
More details on IBM Spyre Accelerator plans will be revealed in 2025,Balakrishnan added.
The Power11
On the processor side, the next generation IBM Power11 system features processor, system, and stack-level innovations to help enterprises propel digital transformation initiatives for their mission-critical infrastructure, Balakrishnan stated.
“IBM Power also continues to support emerging enterprise AI use cases with the MMA (Matrix-Math Assist) architecture,” wrote Balakrishnan.
MMA is a feature of Power10-based servers that handles matrix multiplication operations in hardware, rather than relying solely on software routines. It offers four-times better performance per core for matrix multiplication kernels at the same frequency, according to IBM.
“The Power11 processor is designed to deliver higher clock speeds and can add up to 25% more cores per processor chip than comparable IBM Power10 systems,” Balakrishnan wrote. “The Power11 processor builds upon the key capabilities we delivered with Power10 including stronger reliability, availability and serviceability (RAS) characteristics, better energy efficiency and energy management, and improved quantum-safe security,” Balakrishnan stated.
The Power11 processor will feature a new integrated stack capacitor (ISC) technology and 2.5D packaging along with innovations in cooling such as improved heatsinks and more efficient fans to optimize energy delivery, improve thread and core strength and increase system capacity, according to a blog by William Starke, IBM Power Processor chief architect and distinguished engineer.
In addition, Power11 systems will be based on the recently released DDR5 Dual in-Line Memory Modules and enhanced Open Memory Interface (OMI) interfaces, which enable improved memory reliability, capacity, and bandwidth. Given that OMI is technology agnostic, the Power11 portfolio built on the Power11 processor will also support OMI DDR4 memory migrated from Power10 high-end systems, enabling clients to protect their investments in memory technology, Starke stated.
Power now
IBM has spent a good portion of 2024 developing features to keep the Power servers relevant to enterprise customers in today’s server environment. For example, in October, Big Blue grew its Power Virtual Server infrastructure-as-a-service, which offers Power “pods” of compute, storage, and network infrastructure for private on-premises or IBM Cloud-managed environments. The company added a Power Virtual Server Private Cloud features that lets customers better control and manage Power resources locally or in a hybrid cloud environment, according to IBM.
IBM in May filled out the low end of its Power server portfolio with a 2U rack-mounted server designed for running AI inferencing workloads in remote office or edge locations outside of corporate data centers.
The 1-socket, half-wide, Power10 processor-based system promises a threefold performance increase per core compared to the Power S812 it basically replaces, IBM stated. Running IBM AIX, IBM i, Linux, or VIOS operating systems, the S1012 supports in-core AI inferencing and machine learning with the MMA feature. Each Power S1012 includes four MMAs per core to support AI inferencing.
Developing RPG
The chief application development system widely used by Power servers and its ecosystem is RPG. IBM said it is working with i community developers to come up with AI-based assistants that could help enterprise customers and developers use genAI based code optimized for RPG development.
In 2025, IBM intends to deliver a code assistant for RPG – a generative AI tool which helps developers of IBM i software understand existing RPG code, create new RPG function using natural language description, and automatically generate test cases for RPG code, IBM stated.
Power future lies in chiplets
The last component of IBM’s Power roadmap involved hardware directions after the Power11 server. Specifically, IBM said future Power systems will migrate to a chiplet-based architecture after Power11, according to Starke.
“The idea is to grow effective silicon content per socket, while controlling costs and managing energy and cooling,” Starke stated. “We also discovered that a chiplet-based design will enable improved system topologies, paving the way for continued improvements in large system scaling for multiple generations to come.”
Chiplet designs typically allow for modular design, allowing specific components of a system to be mixed and matched as needed. Intel, AMD, Xilinx and others currently build chiplet systems.
Source:: Network World