• IBM’s Power11 chips are designed for enterprise tasks, focusing on uptime, efficiency, and AI, rather than core count.
• IBM’s Power11 chips target 2x performance per watt over x86 systems.

IBM is rolling out its latest Power11 chips and servers, its first major refresh to the Power line since 2020. The company says the new systems are built to handle modern workloads more efficiently, with features aimed at making AI easier to deploy across business environments.

While competitors like AMD and Intel are pushing high core counts—sometimes over 192 per socket—IBM is taking a different approach. Power11 tops out at 30 cores per socket, using two 15-core dies in one package, similar to the previous Power10 design.

Despite the lower core count, IBM says the individual cores are faster and more efficient. Compared to Power9, the new design offers up to 55% faster performance per core. Power9 was introduced in 2017 and used in US supercomputers like Summit and Sierra, but Power11 isn’t aimed at that space. These systems are built more for sectors like finance, healthcare, and government—industries that rely on stable, data-heavy infrastructure.

Energy and uptime take centre stage

IBM is emphasising power efficiency with this release. The new chips feature a mode that reduces energy use by 28%, trading off a small amount of performance. They also include a new firmware feature called “resource groups” designed to improve system utilisation without increasing energy demands.

Overall, IBM claims Power11 systems deliver twice the performance per watt compared to similar x86 machines. That puts the focus on long-term efficiency rather than raw speed.

The company is also highlighting system reliability. Power11 servers are designed for near-continuous uptime, with an average of just over 30 seconds of unplanned downtime per year. Updates won’t require planned downtime, and core system parts—like fans and power supplies—can be replaced while the machine is running.

To support that level of resiliency, IBM has included extra cores as backup. For example, one system includes 16 cores per chip, but four of those are held in reserve in case one fails.

Built-in security features

IBM says Power11 can respond to a ransomware attack within a minute. That includes detecting threats where attackers encrypt company data and demand payment to unlock it. This is part of the system’s integrated hardware-software design, where IBM controls both layers of the stack.

That tight integration is part of IBM’s strategy. While Nvidia focuses on building and training AI models, IBM is targeting inference—the stage where AI is used to speed up specific business tasks. The company says its goal is to make deploying AI in enterprise settings simpler and more secure.

Hardware layout and models

Power11 will be available in multiple configurations starting July 25, both for on-premise setups and through IBM Cloud.

At the high end is the E1180, which can run as a single-node or four-node cluster. Each node supports up to four sockets, each with 10, 12, or 16 cores. The system supports up to 64 cores and 512 threads per node, plus 16TB of DDR5 memory across 64 DIMM slots. Four cores are kept in reserve as failover.

The E1150 is a similar machine with different core options: 16, 24, or 30 cores per socket, allowing up to 120 total cores per system. It also supports 16TB of memory.

IBM is also offering dual-socket options. The S1124 is a 4U rack-mounted system with up to 60 cores, 8TB of memory, and 24 drive bays. The smaller S1122, a 2U system, offers up to 60 cores, 4TB of memory, and a similar drive layout.

Cost-reduced models, the L1122 and L1124, will only support Linux. That’s a change from IBM’s traditional AIX-focused systems, but likely not an issue for users running Linux-based workloads.

AI add-ons coming later this year

Power11 chips also include built-in AI acceleration, though IBM hasn’t shared full technical details yet. The chips are expected to support running both small and large language models directly on the hardware.

IBM also plans to launch systems with its Spyre AI accelerator later this year. Spyre is a 75-watt add-in card with 128GB of LPDDR5 memory, 200GB/s of bandwidth, and 300 TOPS of AI inference performance. It’s similar in power profile to Nvidia’s L4 card but offers more memory.