# virtio as a device interface

> Your accelerator needs commands and software specific to its chip. Virtio networking lets different makers share a device interface and driver; accelerators still lack that common contract.

Written October 2026. Agenxy — https://agenxy.org/writing/virtio/

---
Your NPU needs the driver for its chip family. Linux's accelerator subsystem
contains six such drivers, with chip-specific commands and, in several cases,
firmware. The [kernel documentation](https://docs.kernel.org/accel/introduction.html)
describes user-space runtimes tailored to the hardware. The shared subsystem
does not supply a shared command set.

Some choices narrow the host too: amdxdna depends on X86_64 and AMD_IOMMU.
If the runtime stops being maintained, keeping the chip need not keep it useful.
For network devices, virtio offers another arrangement: one interface and
drivers shared across makers.

## One driver can serve network devices from different makers

[virtio](https://docs.oasis-open.org/virtio/virtio/v1.4/virtio-v1.4.html)
specifies registers, queues and features for each defined device type. A
conforming network device presents that interface regardless of its maker.
The specification was designed for virtual environments, where the devices
look physical to the guest.

Its table lists 47 device numbers, 23 with full definitions. A conformance
claim covers the base clause, one transport (PCI, MMIO or Channel I/O) and a
defined device type. Linux and
[FreeBSD](https://github.com/freebsd/freebsd-src/tree/main/sys/dev/virtio)
include drivers. [virtio-win](https://github.com/virtio-win/kvm-guest-drivers-windows)
develops Windows drivers openly under a BSD licence; Microsoft does not ship
them. Apple's [Virtualization framework](https://developer.apple.com/documentation/virtualization)
also supplies virtio devices to virtual machines.

Hardware can present the interface. Linux's
[vDPA drivers](https://elixir.bootlin.com/linux/latest/source/drivers/vdpa/Kconfig)
put the data path into several makers' hardware, with vendor-specific control
alongside it. NVIDIA's [BlueField guide](https://docs.nvidia.com/doca/sdk/doca+virtio-net+service+guide/index.html)
says its virtio-net devices need no QEMU, emulator or vendor driver in the
guest. Vendor software remains on the host or card. Inside the guest, the
existing network driver is enough.

## Open on its own terms

The specification is public HTML and PDF without a login. It permits copying,
publication and explanatory works, but not modification of the document itself.
Under the [three tests](@what-open-means-for-hardware), Read passes.

The [OASIS Non-Assertion Mode](https://www.oasis-open.org/policies-guidelines/ipr/)
extends the essential-claims promise to third-party implementers, without
requiring membership or a fee. Participants are bound, not every patent
holder. A party may suspend its promise against a beneficiary asserting an
essential claim over an implementation of the same specification. Red Hat's
[disclosures](https://github.com/oasis-tcs/virtio-admin/blob/master/IPR.md)
commit claims to those terms if they prove essential.

The specification has no logo or certification programme. OASIS reserves its
name and specification names for references to its work while welcoming
implementation and use ([trademark policy](https://www.oasis-open.org/policies-guidelines/trademark/)).
The terms examined do not bar an implementer from describing conformance.

The transport adds another question. PCI relies on
[PCI-SIG's specifications](@pcie-and-m2-licence-terms), paid for by
non-members, and requires vendor ID 0x1AF4. Linux and
[QEMU](https://www.qemu.org/docs/master/specs/pci-ids.html) attribute that ID
to Red Hat, which donated part of its range for virtual devices. Permission
for another hardware maker to use it was not established.
[BlueField documentation](https://docs.nvidia.com/doca/archive/2-5-2/NVIDIA+DOCA+Emulated+Devices)
shows 1af4:1041. The MMIO text is modelled on PCI, but no PCI-SIG dependency
was found there.

The interface's grant reaches non-members. Its PCI identifier still needs
an answer from its holder.

## Virtio has no accelerator device

The specification defines no NPU device. Its accelerator reference concerns
cryptography. Searches of the
[committee repository](https://github.com/oasis-tcs/virtio-spec) found no
NPU, inference, tensor or neural-device proposal. A GPU-computing request for
AMD ROCm uses the existing GPU device.

Outside the committee, [AMD's prototype](https://github.com/amd/virtio-npu)
exposes an NPU through QEMU's virtio-accel-pci device and AMD's driver. It
reuses GPU device ID 16 with AMD-specific capabilities. Redistribution is
permitted in unmodified binary form only.

The independent [virtio-accel project](https://github.com/MicroPerceptron/virtio-accel)
describes transport-neutral, vendor-agnostic devices with Apple, Intel, AMD
and Qualcomm back ends. It claims no device ID. None of those other vendors
was shown to participate. Neither prototype has its own reserved device
number or a completed device clause in the specification.

A reservation can precede a definition. The recent numbers 50 and 51 cover a
video codec and inter-VM messaging; only 23 of 47 listed numbers have clauses.
A prototype has not yet supplied the common contract.

## Two devices, two futures

Linux shows both arrangements: maker-specific accelerator commands and a
network interface shared across implementations. Rockchip's driver is even
maintained apart from its maker, with its user-space part in Mesa. Publishing
an individual interface still does not make it the common device type.

The NPU owner may need the kernel driver, firmware and user-space runtime.
An open kernel component does not establish freedom to change the rest.
Virtio networking shows how one layer can stop depending on one maker's stack.

## A driver for every accelerator

Agenxy proposes an OASIS accelerator device type under Non-Assertion Mode:
a number, a device clause and implementations. Its common contract must name
what different accelerators can share. A standard hardware-facing device
could then use a common operating-system driver, while firmware and card-side
software might remain, as on BlueField.

The PCI identifier also needs a written grant for conforming hardware from
its holder. A shared driver will not settle firmware rights, but it can
remove one dependency on the chip maker's software.

The owner should be able to keep the accelerator useful. The work is to
specify the device, rather than require a new stack for every chip.

## Sources

Linux drivers and interface headers were examined on 14 October 2026. The six
families were AMD, Arm, Habana, Intel, Qualcomm and Rockchip; examples include
DRM_IOCTL_AMDXDNA_CREATE_HWCTX and DRM_IOCTL_IVPU_BO_CREATE. Virtio 1.4 is a
Committee Specification dated 8 April 2026. OASIS IPR section 10.3 supplies the
non-assertion terms. Red Hat filed 42 statements naming 158 US patents and 36 published
applications. The GPU-computing request is from 2025; BlueField's ID example
is in DOCA 2.5.2.

The defined types include network, block, console, entropy, GPU, input, sound,
file system, I2C, GPIO, SPI, CAN and RTC. Apple's framework also provides memory
balloon and socket devices. vDPA hardware includes Intel, NVIDIA,
AMD/Pensando, Marvell, SolidRun and Alibaba.

FreeBSD uses a BSD licence. virtio-win's licence begins with Red Hat copyright;
Fedora and Red Hat distribute binaries, cross-signed for Fedora and signed
through Microsoft WHQL in RHEL.
Remaining kernel sources:

- [drivers/accel](https://elixir.bootlin.com/linux/latest/source/drivers/accel).
- [include/uapi/drm](https://elixir.bootlin.com/linux/latest/source/include/uapi/drm).

Other sources were consulted on 15 October 2026.