ppc64le
The name Linux, Red Hat packages, container images and Python wheels use for a 64-bit IBM Power processor running in little endian byte order, and the word an IBM i shop has to match when it installs AI software on Power hardware. IBM i itself is not ppc64le, so a ppc64le build runs in a Linux partition on the same Power server, never inside IBM i or PASE.
What does ppc64le mean?
Every computer has to decide which end of a number to store first. That choice is called byte order, or endianness. Think about writing a date. 30-10-2026 and 2026-10-30 mean the same day, but the pieces come in opposite order. A reader who expects one and gets the other reads it wrong. Chips and compiled programs have the same trouble with numbers.
IBM Power chips can work either way, and the operating system picks one and keeps it. The "le" at the end of ppc64le means little endian, the same order x86_64 uses, which is the order nearly all Linux software is built for now. The older big endian build of Linux on Power is plain ppc64. IBM is blunt about mixing the two. Its little endian FAQ says "the operating system enablement only supports applications of the same type. As such, a little endian operating system (ppc64le or ppc64el) can only run little endian applications built for this software platform."
One architecture, four names you will actually type
| Where you meet it | The name | A real example |
|---|---|---|
| Red Hat package files and repository paths | ppc64le | Red Hat's 7.1 notes give the file gcc-4.8.3-9.ael7b.ppc64le.rpm |
| Debian and Ubuntu | ppc64el | IBM's own FAQ writes the pair as "ppc64le or ppc64el" |
| Container images | linux/ppc64le | Docker Hub lists linux/ppc64le among the platforms for ubuntu:latest |
| Python wheel file names | manylinux_2_34_ppc64le | torch-2.6.0+ppc64le1-cp312-cp312-manylinux_2_34_ppc64le.whl on IBM's Power wheel index |
A wheel is a prebuilt Python package file. Its platform tag format comes from PEP 600, the Python packaging standard for portable Linux wheels, which is marked Final and defines manylinux_${GLIBCMAJOR}_${GLIBCMINOR}_${ARCH}. PEP stands for Python Enhancement Proposal, the way a Python standard is written down. The number in the middle is the oldest glibc the file will run against, and glibc is the main C library on Linux. Red Hat, Docker Hub, IBM and Python pages read October 30, 2026.
When did Power get a little endian Linux?
Red Hat added the little endian build in Red Hat Enterprise Linux 7.1, whose notes say "Only IBM POWER8 processor-based servers are supported with Red Hat Enterprise Linux for POWER, little endian" and that before that release only the big endian version was offered for IBM Power Systems. The floor later moved up. The Red Hat Enterprise Linux 10 notes list "IBM Power Systems, Little Endian (POWER10 and later)", so an older Power 9 box is off that list.
Little endian on Power, and how long it took to reach AI tooling
- October 2014IBM ships PowerKVM 2.1.1. IBM's little endian FAQ names it as a way to run a little endian Linux guest on a POWER8 machine.
- April 26, 2015Debian 8, called jessie, is released. Its architecture list includes POWER processors, which Debian spells ppc64el.
- March 31, 2026IBM's Power wheel index publishes PyTorch 2.6.0+ppc64le1 for Python 3.11 and 3.12, two files of about 116 MB each.
- October 30, 2026We read PyPI. The newest PyTorch, NumPy, SciPy, ONNX Runtime and vLLM releases carry 76 Linux wheels between them, and not one is ppc64le.
PowerKVM date from IBM's support page "More FAQs about Little Endian: An Update". Debian date and architecture list from the Debian 8 release page. PyTorch wheel names, sizes and upload times from IBM's wheel index page for torch 2.6.0+ppc64le1. PyPI counts taken from the PyPI package records on October 30, 2026.
Can IBM i run ppc64le software?
No. IBM i is not a ppc64le operating system, and neither is AIX. They sit on the same Power chips, in the other byte order, with a different program format. Three things have to line up before a compiled file will run, and for ppc64le on IBM i none of them do.
- Byte order. The Mono project's 2018 write-up on porting to AIX and IBM i states it plainly: "Both of our targets run on 64-bit PowerPC processors in big endian mode."
- File format. The same write-up says "While most Unix-like operating systems use ELF, AIX (and by extension, IBM i for the purposes of this port) use XCOFF, a subset of the Windows PE binary format." ELF and XCOFF are two different ways of wrapping up a compiled program.
- The runtime. IBM's IBM i 7.6 documentation calls PASE for i "an integrated runtime environment for AIX applications running on the IBM i operating system", and says its programs "Use the same binary executable format as AIX PowerPC applications". PASE stands for Portable Application Solutions Environment. It is an AIX runtime, not a Linux one.
- So a ppc64le wheel or container is not something PASE can open. It was built for a different operating system, in the opposite byte order, in a format IBM i does not load.
The hardware is still shared, and that is the practical answer. One Power server can run IBM i in one partition and a little endian Linux in another. IBM's FAQ says the isolation that separates operating system levels "applies to big and little endian operating systems with Linux and other Power operating systems such as AIX and IBM I". That is how an IBM i shop serves a model next to production: a Linux partition on the box it already owns, or a Linux instance in IBM Power Virtual Server.
Why do AI installs fail on a Power box?
pip, the Python installer, looks for a prebuilt file that matches your Python version and your platform. When nothing matches, pip either compiles the package from source, which can take hours, or stops and says there is no matching distribution. Most AI projects build those files for two platforms only: Intel and AMD chips, which Python calls x86_64, and 64-bit Arm, which it calls aarch64. Power is not on the list.
Prebuilt Linux wheels on PyPI for the newest PyTorch, NumPy, SciPy, ONNX Runtime and vLLM releases, read October 30, 2026
Counted from the PyPI package records for numpy 2.5.4, scipy 1.18.1, torch 2.15.0, onnxruntime 1.31.1 and vllm 0.32.0 on October 30, 2026. Only Linux wheels are counted, so Windows and macOS files are left out. PyPI is the Python Package Index, the place pip downloads from by default.
Packages with no compiled code are fine. transformers 5.22.1 ships one wheel named py3-none-any, which means pure Python and any platform, so it installs on Power with no special help. The trouble is in packages with compiled parts underneath, and those are most of the fast ones.
Where do you get ppc64le AI builds?
Three ways to get AI software onto a Power Linux partition
- Use IBM's Power wheel indexIBM runs a wheel repository for this architecture and describes it as "a curated repository of natively built Python wheels for IBM Power (ppc64le), optimized for AI, data science, and scientific computing". The install line IBM publishes is an ordinary pip install of one pinned package version, with the prefer binary flag set and the extra index URL flag pointed at https://wheels.developerfirst.ibm.com/ppc64le/linux, so pip checks IBM's Power files as well as PyPI. IBM's documentation lists Power 10 and Power 11, Python 3.11 and 3.12, and names PyTorch, TensorFlow in select versions, NumPy, SciPy, pandas, scikit-learn and OpenBLAS.
- Pull a container image built for linux/ppc64leRegistries publish one image per platform under the same name. Docker Hub lists linux/ppc64le among the platforms for ubuntu:latest, so the base layer is there. Check the platform list of the AI image you want before you build a plan on it, because many are x86_64 only.
- Build it from source on the machineThis always works and always costs time. You need the compilers, the dependencies and the patience, and the result is yours to rebuild at every upgrade. Try the first two before you take this one.
IBM's wheel repository page is published under IBM Open-Source AI Foundation for Power 2026.06.0, and the versioned snapshot path reads linux-v2026.06.0. The install command is from an IBM Community post by Nikhil Kalbande dated August 1, 2025. Pages read October 30, 2026.
Expect the Power build to trail the headline release. On October 30, 2026 the newest PyTorch on PyPI was 2.15.0 with no Power file at all, while IBM's Power index topped out at 2.14.0, marked "Latest version on stage". One release behind is close, not current. Check the exact version your project needs before you promise a date, and remember Red Hat Enterprise Linux 10 supports Power little endian only on "POWER10 and later".
Related Terms
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.
IBM PASE for i
PASE for i is the AIX runtime built into IBM i, and it is where the open-source software behind most AI work on the platform actually runs. It ships as option 33 of 5770-SS1, has been a no-charge option since V5R2M0, and it hosts the yum-installed Node.js and Python builds under /QOpenSys/pkgs, including the Mapepire server that IBM's IBM i MCP Server connects through on port 8076.
IBM AIX
AIX is IBM's UNIX operating system for IBM Power hardware, the sibling of IBM i on the same servers and the source of the AIX runtime that IBM i ships as PASE for i. On Power 11 it needs AIX 7300-03-01 or higher to run in Power11 compatibility mode, and its AI packages arrive as RPMs from the AIX Toolbox for Open Source Software.
IBM Power 11
IBM's current generation of Power processor and server hardware, announced July 8, 2025 and generally available July 25, 2025, carrying forward Power10's on-chip Matrix Math Acceleration and becoming the first Power generation to support the IBM Spyre Accelerator card.
IBM Power 10
IBM's Power processor generation that preceded Power11, announced August 17, 2020 and shipping from September 2021, the first Power chip to embed Matrix Math Acceleration into every core for on-chip AI inferencing without a separate GPU. IBM began winding down new Power10 sales in February 2026.
IBM Power 9
Power 9 is the IBM Power processor generation before Power 10, available from 20 March 2018 in the scale-out S914, S922 and S924 models that IBM i supports. For AI it falls one generation short: on-chip Matrix Math Acceleration starts with Power 10, the Spyre accelerator card is a Power 11 option, and IBM states that IBM i 7.6 is not supported on Power9 and earlier processor technology, which caps a Power 9 at IBM i 7.5.
vLLM
An open source server that keeps a large language model loaded and answers requests over an OpenAI style API, with PowerPC CPUs on its supported hardware list. That lets an IBM i shop run a model in a Linux partition on the same IBM Power server as its production workload, and Red Hat ships a generally available vLLM container image for ppc64le in Red Hat AI Inference 3.5.0.
Red Hat OpenShift AI
Red Hat's add-on platform for serving and governing AI models on a Red Hat OpenShift cluster, supported self-managed on ppc64le, the IBM Power architecture that also runs IBM i. It lets a Power shop host open weight models on its own hardware instead of calling an outside service.
IBM Power Virtual Server (PowerVS)
IBM Power Virtual Server, or PowerVS, lets an IBM i shop rent IBM Power server capacity inside IBM's data centers instead of buying a box, and run the same AIX, IBM i and Linux partitions there. IBM describes it as a family of configurable, multitenant, virtual IBM Power servers with access to IBM Cloud services, web, DevOps and IBM Db2, which is how an IBM i partition ends up on the same cloud network as IBM Cloud AI services such as watsonx.ai.
IBM Spyre Accelerator
A PCIe4 AI inferencing card IBM ships for IBM Z, LinuxONE and Power11 systems to offload generative and agentic AI workloads from the main processor. On Power11 it is driven from a Red Hat Enterprise Linux or Red Hat OpenShift AI partition, and as of September 2026 IBM has not documented a way to use it from an IBM i partition.