Dynamic Logical Partitioning
Dynamic logical partitioning, or DLPAR, moves processors, memory and card slots in and out of a running IBM Power partition without restarting it. IBM documents it for the Spyre Accelerator, so an AI card can join or leave a Linux inference partition while the IBM i production partition keeps working.
What does it actually do?
A Power server can be split into separate partitions. Each one runs its own operating system, so IBM i, Linux and AIX can sit on the same machine. Dynamic logical partitioning, or DLPAR for short, is how you change what a partition owns while it is still running. Think of a partition as a desk and the resources as chairs. DLPAR lets you carry a chair to another desk without asking anyone to stand up and leave the room.
- Processors can be added, moved or removed
- Memory can be added, moved or removed
- Input/output slots can be added, moved or removed, and a slot is where a card like Spyre lives
Why it matters for a Spyre card
The Spyre Accelerator is the AI inference card IBM sells for Power 11. It plugs into a PCIe slot, short for Peripheral Component Interconnect Express, inside an ENZ0 expansion drawer, and it is driven from a Linux partition rather than from IBM i. For a shop with one server, that raises a fair question: to change which partition holds the cards, does the whole box have to come down? IBM answers it in its own Spyre publication. Spyre devices can be added to or removed from a logical partition dynamically, the same as other PCI devices on a Power system. So a card can start the month in a test partition and finish it in a production one, and the IBM i partition next door never stops.
What IBM says to do before and after
The two practices IBM documents for moving a Spyre card
- Before you remove a cardStop every container that is using that Spyre accelerator. Pulling the device out from under running work is not supported.
- After you add a cardRerun the servicereport tool and restart the active Spyre containers. IBM says this checks that the resource limits suit the new card count and prevents out of memory problems.
Spyre support for these operations arrived with the IBM Spyre Enablement Stack for Power 1.1, delivered on Red Hat AI Inference Server 3.4 in March 2026.
How many Spyre cards fit
Maximum Spyre cards per system
One logical partition supports up to 24 cards. A drawer holding 12 cards needs Red Hat AI Inference Server 3.5 or later. Figures are IBM's own, from its Spyre Accelerator for Power publication.
Minimum resources IBM lists for a Spyre partition
| Resource | Minimum |
|---|---|
| Memory | 512 GB |
| Processor cores | 10 |
| Storage | 600 GB |
IBM calls the memory figure a minimum per model and says to raise it as models get larger and more copies run at once.
What can stop it
IBM lists one case where a partition does need a reboot anyway. Running three four-card containers for retrieval augmented generation, the technique of looking up your own documents before the model answers, in a single 12-card drawer with a large input context and a large batch size can push the cards into a failed state. IBM's guidance is to run no more than two of those containers per drawer and give the spare cards lighter work such as embedding, reranking or entity extraction.
Related Terms
AI Inference Partition
A separate supported operating-system partition used to execute an AI model while an IBM i partition may remain the production application and data source.
Spyre Accelerator
IBM's PCIe AI inference card. On IBM Power 11 it carries 32 cores, 128 GB of memory, and more than 300 TOPS inside a 75 watt envelope.
ENZ0 Expansion Drawer
The PCIe4 expansion drawer IBM documents for supported Spyre Accelerator for Power configurations on Power 11 systems.
AI Workload Placement
The decision about which system, operating environment, and partition executes an AI model or agent, separately from the IBM i application supplying its data.