GPU and PCIe passthrough

AtomOS exposes physical PCIe devices to guests when the host firmware, IOMMU, and device grouping allow safe assignment. The common case is whole-GPU passthrough (one or more cards dedicated to a VM). Partitioning technologies (vGPU, MIG, SR-IOV) are separate capabilities—confirm per release before planning capacity.

PCI passthrough overview

Passthrough gives a guest direct access to hardware (typically NVIDIA/AMD GPUs, sometimes other accelerators). The hypervisor binds the device via VFIO after IOMMU isolation.

flowchart TB
  HW[Physical GPU / PCIe device]
  IOMMU[VT-d / AMD-Vi IOMMU groups]
  VFIO[VFIO driver binding]
  VM[KVM guest with native driver]
  HW --> IOMMU --> VFIO --> VM

Host prerequisites

Requirement	Action
CPU virtualization	VT-x / AMD-V enabled in firmware
IOMMU	Intel VT-d or AMD AMD-Vi enabled
Kernel	IOMMU boot parameters per distro docs (validate on AlmaLinux 9 host)
Device visibility	sudo atomos-cli --pci-list and lspci -nn on the host
Isolation	All functions in the IOMMU group that must move together (GPU + audio function on same card)

Validate before purchase

Record vendor:device IDs from lspci -nn (for example 10de:24b0) for every GPU you plan to pass through. Match them in the VM spec.

IOMMU groups and safety

• Devices in the same IOMMU group cannot be isolated from one another for DMA purposes.
• Passing through one function often requires passing related functions or using ACS-capable topologies.
• Do not assume a cloud or nested-virt instance exposes usable IOMMU for passthrough—validate on bare metal for production GPU workloads (Hardware).

Guest configuration

1. Create a VM with adequate CPU, memory, and NUMA awareness for GPU workloads.
2. Request devices in the pledge/spec using PCI identifiers (see example below).
3. Install vendor guest drivers (NVIDIA, AMD, etc.) inside the VM.
4. Run representative AI/ML, CUDA, or graphics tests before production cutover.

Example request fragment

{
  "pci": [{ "vendor": "10de", "model": "24b0", "quantity": 1 }]
}

Field	Meaning
vendor	PCI vendor ID (hex, without 0x)
model	PCI device ID (hex)
quantity	Number of matching devices to assign

Use IDs from lspci or atomos-cli --pci-list on the target host—IDs differ by SKU and slot.

Migration, HA, and maintenance limits

Topic	Typical constraint
Live migration	Not supported for guests with assigned physical GPUs—plan stop or recreate on another GPU host
Host failure recovery	Recreate requires free GPU with matching IDs on a peer (Availability)
Planned maintenance	Shutdown guest; drain GPU host like any non-migratable workload
Driver updates	Host and guest driver upgrades may need coordinated reboots

Document GPU-backed VMs explicitly in runbooks (Availability & recovery).

GPU partitioning: vGPU, MIG, SR-IOV (distinct from passthrough)

These technologies slice or virtualize accelerator capacity. They are not the same as whole-device passthrough and must be checked per AtomOS release and hardware generation.

Technology	Idea	Documentation status
NVIDIA vGPU	Licensed virtual GPU profiles on supported datacenter cards	Planned / unverified in public AtomOS docs—not GA unless confirmed for your build
NVIDIA MIG	Hardware partitions on A100/H100 class GPUs	Planned / unverified — confirm with Elemento
SR-IOV	Virtual functions on NICs or some accelerators	Planned / unverified for general GPU SR-IOV — confirm per device and release
AMD MxGPU	AMD virtualized GPU	Planned / unverified — confirm with Elemento

Do not assume GA

If a sales conversation or lab demo mentioned vGPU or MIG, still verify package availability, licensing, and Electros specfile fields on the exact release you deploy. Until verified, plan whole-GPU passthrough only.

Recommended validation path

1. Confirm GPU visibility on the host; record vendor and device IDs.
2. Enable IOMMU; review PCIe groups; isolate required functions.
3. Create a test VM with the PCI fragment and install guest drivers.
4. Benchmark a representative workload.
5. Record migration and HA exclusions in your operational docs.
6. If you need shared GPU capacity across many VMs, open a capacity planning discussion with Elemento before relying on partitioning features.

Related reading

• Hardware requirements — IOMMU and sizing
• Confidential computing (TDX) — distinct trust-boundary workloads
• Use cases — private AI