ppc64le
ppc64le is the architecture code for 64-bit little endian Linux on IBM Power, and it is the only kind of build that will run on a Power server. Every part of an AI stack, from the serving engine to the Spyre Accelerator drivers, has to exist as a ppc64le build before an IBM i shop can serve a model in a Linux partition on the same machine.
What is ppc64le?
A processor can store the bytes of a number in two orders. Big endian puts the biggest byte first. Little endian puts the smallest byte first. Think of writing a date. 27-10-2026 and 2026-10-27 hold the same facts, but a program that expects one order and reads the other gets nonsense.
IBM says the Power architecture is bi-endian, meaning the chip itself can read data both ways. The operating system has to pick one and stay with it. ppc64le is the name for the little endian side: 64-bit Linux on Power, in the same byte order Linux uses on an Intel or AMD server. That matching order is the whole reason the mode exists. IBM's own explanation is that clients, software partners and IBM developers said porting their software to Power becomes simpler when the Linux environment on Power supports little endian mode.
Why will an x86 AI container not run on a Power server?
Because the program inside it was compiled for a different instruction set and a different byte order, and nothing translates it while it runs. IBM puts it plainly: a little endian operating system, written ppc64le or ppc64el, can only run little endian applications built for that software platform.
So every layer of an AI stack needs its own Power build. IBM keeps open source container images for Power in the IBM Container Registry under icr.io/ppc64le-oss/, and says IBM builds them itself in a secure environment so the images are known to be genuine. That catalog runs to well over a hundred images and includes vLLM, Ollama and Milvus. vLLM stands for virtual large language model, and it is the engine that serves prompts and streams answers back.
A Power build existing is the first question, not a formality. Red Hat's own Spyre serving image shows how uneven this can get: the same vllm-spyre-rhel9:3.2.5 image is Generally Available on ppc64le and s390x, while the x86 variant is listed as Technology Preview.
Does IBM i or AIX have to change?
No. IBM answers this one directly: there is no effect on AIX or IBM i application environments as a result of the change. Both stay big endian. Nothing in your production workload moves.
ppc64le matters because the AI work runs beside IBM i, not inside it. On a Power server you give the AI job its own Linux partition, and that partition is the little endian one. It is the partition that sees a Spyre Accelerator card and serves the model, while IBM i keeps running in its own partition on the same box.
Which AI runtime parts have a ppc64le build?
Enough of them that a supported Power inference stack exists today, and the vendors name the architecture in writing rather than leaving you to guess.
What Red Hat and IBM publish for ppc64le on IBM Power
| Part | ppc64le status | What the vendor says |
|---|---|---|
| Red Hat AI Inference Server 3.2.5 | Named architecture | Container images with multi-architecture support for s390x on IBM Z and ppc64le on IBM Power |
| vllm-spyre-rhel9:3.2.5 image | Generally Available | GA on s390x and ppc64le; the x86 variant is Technology Preview |
| vLLM serving engine | Packaged for Power | Red Hat AI Inference Server 3.2.5 packages the upstream vLLM v0.11.2 release |
| IBM Spyre Accelerator | Generally Available | Red Hat AI 3 lists the accelerator as IBM Spyre for Power (ppc64le) |
| vllm-spyre-rhel9:3.5.0 cluster image | Published as a set | Listed with IBM Spyre Enablement Stack 1.1.1 and vllm-spyre 1.6.1 |
| Open source AI containers | Prebuilt by IBM | icr.io/ppc64le-oss/ in the IBM Container Registry, including vLLM, Ollama and Milvus |
Red Hat supports single host Spyre Accelerator deployments on IBM Power and IBM Z on RHEL AI 9.6 and later. Sources: Red Hat AI Inference Server 3.2 release notes, the Red Hat AI 3 supported product and hardware configurations guide, and IBM's announcement of Power open source images in the IBM Container Registry.
Three checks before you plan an AI tool for a Power server
- 1. Find the architecture named in the vendor's support listNot in a blog post, in the support list. Red Hat names IBM Spyre for Power (ppc64le) in its supported configurations, and names ppc64le among the architectures its inference server images cover. A project that only ships x86 images and x86 Python wheels has nothing for you to install.
- 2. Match the Linux versionRed Hat supports single host Spyre deployments on RHEL AI 9.6 and later. Red Hat Enterprise Linux 9.6 became generally available on 20 May 2025, so an older partition is not a candidate until it is upgraded.
- 3. Take the versions as a setRed Hat publishes the vllm-spyre-rhel9:3.5.0 cluster image together with IBM Spyre Enablement Stack 1.1.1 and vllm-spyre 1.6.1. Those numbers are listed as a pairing, so treat them as one unit instead of picking parts.
Version numbers move. Confirm the current pairing in Red Hat's supported configurations before you order hardware.
When did Linux on Power become little endian?
From the first little endian Power chip to a supported AI serving stack
- 23 April 2014IBM unveils POWER8 at an Open Innovation Summit with the OpenPOWER Foundation in San Francisco. Little endian mode arrives with this generation. The first scale out servers, including the Linux only S812L and S822L, became available on 10 June 2014.
- 5 March 2015Red Hat Enterprise Linux 7.1 ships. Red Hat writes that it introduces little endian support on IBM Power Systems servers using IBM POWER8 processors, and that packages built for it carry the ppc64le architecture code.
- 20 May 2025Red Hat Enterprise Linux 9.6 becomes generally available. This is the floor Red Hat names for single host Spyre Accelerator deployments on Power.
- 11 December 2025IBM announces the general availability of Red Hat AI Inference Server on IBM Power, putting a supported model serving stack on the ppc64le side of the machine.
Red Hat's 7.1 notes also carried a limit that has long since lifted: at launch the little endian build was supported only as a KVM guest.
ppc64le tells you which builds exist, not how fast they run. It answers whether a tool can be installed on your Power server at all. How many users one Spyre Accelerator card then supports is a separate question, settled by the serving engine settings, the model and the memory in the partition.
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.
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.
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.
vLLM
The open source model serving engine inside Red Hat AI Inference Server, known for PagedAttention and continuous batching.
IBM i
IBM's operating system for Power servers running traditional AS/400-lineage business applications, with Db2 for i built directly into the OS.