AMD Instinct MI300/MI350 GPU Spatial Partitioning (XCP) Helper
A small DRM helper that supports AMD's spatial partitioning (XCP) feature on Instinct data-center accelerators such as the MI300X, MI325X, and MI350X/MI355X, allowing a single GPU to be carved into multiple isolated compute slices for AI and HPC workloads. It works alongside the main amdgpu driver rather than driving hardware on its own.
recommendation
It should stay because the underlying hardware is brand new — the MI350X launched in June 2025 — and AMD actively documents partitioning across its current Instinct lineup. Upstream activity confirms ongoing maintenance, including a 2025 fix for memory release on partition teardown, and there is no alternative implementation outside amdgpu itself. Deployments are narrow (data-center AI/HPC) but the code is current and necessary for that use case.
repository signals
sources
- git.kernel.org
Initial upstream addition says amdxcp was added to support amdgpu spatial partition.
- git.kernel.org
Upstream saw a real functional fix in 2025 for unplug/release of XCP DRM memory, indicating current maintenance rather than abandonment.
- instinct.docs.amd.com
AMD documents MI300X compute partitioning modes implemented in the driver layer, matching the XCP/spatial-partition use case.
- instinct.docs.amd.com
AMD virtualization documentation lists spatial partitioning support for Instinct MI300X, MI325X and MI350X/MI355X, showing the feature spans current products.
- amd.com
AMD product page shows Instinct MI350X launched on 2025-06-12, so relevant hardware was still being sold new in 2025.
codex reasoning notes (technical)
Local shell inspection shows this is a real DRM helper module, not headers/tests. Local `git -c safe.directory=... log/show` found 7 substantive commits from 2023-06-09 through 2025-08-04, including a 2025 functional bugfix, so upstream attention is active. A shell `lei` lore query could not run because `lei` is not installed; web lore searches returned no public removal-thread hits, so I found no evidence of deprecation/removal discussion. URLs for the two kernel commits are canonical kernel.org commit pages constructed from hashes obtained via shell; AMD deployment/product URLs were obtained via web search. Hardware scope is niche data-center AI/HPC partitioning rather than broad consumer deployment, so deployments are low, but the hardware/feature set is current and has no natural replacement outside amdgpu itself.