IBM Spyre Enablement Stack for Power
The drivers, runtime and models IBM requires before a Spyre Accelerator card runs anything on a Power 11 server. It arrives from Red Hat inside Red Hat AI Inference Server, so the card sits idle until the Linux partition that drives it has the stack installed and entitled.
What does it do?
A Spyre Accelerator card is a piece of hardware sitting in a drawer. On its own it does nothing. The enablement stack is the set of parts that lets a Linux partition on the same Power 11 server find the card, load a model onto it, and answer requests. IBM lists those parts as drivers, vLLM backends, runtimes and models. vLLM stands for virtual large language model, and it is the serving engine that hands a prompt to the card and streams the answer back.
- Drivers, so Red Hat Enterprise Linux can see the card at all.
- A vLLM backend and runtime built for ppc64le, the Power build of Linux.
- Models that are already prepared for the card, so nothing has to be compiled first.
- Zero copy access from the Power processor straight to the card, which IBM calls VFIO.
- Tools for watching performance and finding faults, including the aiu-smi tool.
This is a Linux side stack. IBM requires Red Hat AI Inference Server or Red Hat OpenShift AI, so the partition that drives the cards is a Red Hat Enterprise Linux partition on the same physical server, not an IBM i partition.
What do you have to order?
Four program numbers appear on the IBM prerequisite list, and the stack needs a live Red Hat subscription for both the operating system and the inference server.
- 5639-SPY ... IBM Spyre Enablement Stack for Power
- 5639-RIS ... Red Hat AI Inference Server
- 5639-SPA ... IBM Spyre Stack Updates for Power
- 5639-SAI ... IBM Open-Source AI Foundation for Power
IBM also expects system firmware at platform level, root or sudo access, and outbound network access to download models unless you use the offline bundle.
How much of the partition does it need?
IBM minimum resources for the partition that drives the cards
| Resource | Minimum |
|---|---|
| Memory | 512 GB |
| Processor cores | 10 |
| Storage | 600 GB |
IBM says to read the memory figure as a minimum per model and to raise it as the model size and the number of pods grow.
How do you install it?
IBM install order on a Power 11 Linux partition
- Check the prerequisitesRed Hat Enterprise Linux 9.6, 9.8 or 10.2, registered with Red Hat, an active Red Hat AI Inference Server subscription, and system firmware at platform level.
- Install the Linux packagesOne IBM command installs container-tools, git, git-lfs and wget, because the inference server is delivered as a container image.
- Add the IBM Power toolsConfigure the IBM Power Tools repository and install ServiceReport, then ibm-power-managed for a system managed by a Hardware Management Console or ibm-power-nonmanaged for a stand-alone system.
- Start the inference serverSet up the environment and the model directory, then run Red Hat AI Inference Server either as a stand-alone Podman container or as a Podman Quadlet that comes up with the partition.
Order follows the configuration and installation chapter of the IBM Spyre Accelerator for Power documentation at release 1.3.
What changed in each release?
IBM Spyre Enablement Stack for Power releases
- December 2025Release 1.0, with Red Hat AI Inference Server 3.3. The first release covered the product overview, the requirements, the configuration and the serviceability features.
- March 2026Release 1.1, with 3.4. Added chunked prefill, prefix caching, dynamic logical partitioning of Spyre cards, and Red Hat OpenShift AI.
- June 2026Release 1.2, with 3.5. Added 12 card configurations, custom fine tuned Llama weights, and better error recovery.
- September 2026Release 1.3, with 3.6. Added Granite 4.1, Granite vision and Mistral models, and a new metric in the aiu-smi tool.
Drawers with 12 cards need Red Hat AI Inference Server 3.5 or later.
Release 1.1 shortened the wait a person actually feels. Before chunked prefill, the card read the whole prompt before writing the first word of the answer, which IBM measured at 15 to 20 seconds on a large request. Chunked prefill breaks that reading into pieces, so the card can switch between reading and writing and start answering sooner.
Wait before the answer starts, on a large request
IBM figures. Chunked prefill is on by default with a chunk size of 1024 tokens, and it applies to decoder models, not to embedding or reranker models.
Related Terms
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.
Red Hat AI Inference Server
Red Hat's model serving product, built on the open source vLLM engine. IBM lists it as one of two supported ways to drive a Spyre Accelerator on IBM Power.
vLLM
The open source model serving engine inside Red Hat AI Inference Server, known for PagedAttention and continuous batching.
ENZ0 Expansion Drawer
The PCIe4 expansion drawer IBM documents for supported Spyre Accelerator for Power configurations on Power 11 systems.
Red Hat OpenShift AI
Red Hat's larger AI platform, and the second option IBM accepts for driving a Spyre Accelerator on IBM Power.