OPEN SOURCE

NOVA Microhypervisor

Trusted isolation for shared AI infrastructure.

Open-source microhypervisor with hardware-enforced isolation across workloads, devices, and memory. Now with DMA remapping support on AMD platforms using IOMMU virtualization.

256 TB physical memory

128 PB virtual address space per workload

GPLv2 open source

WHY THIS MATTERS

AI infrastructure changes the requirements for infrastructure isolation.

Traditional infrastructure architectures were built around predictable enterprise applications and bounded execution models. AI workloads increasingly behave differently.

Modern AI systems introduce:

  • continuous execution

  • dynamic orchestration

  • shared accelerators

  • multi-tenant infrastructure

  • autonomous runtime behavior

As infrastructure utilization increases, isolation boundaries become increasingly important — particularly around memory access, device access, and shared execution environments.

NOVA was designed around a simple principle:

Critical isolation guarantees should be enforced at the hardware boundary whenever possible.

WHAT NOVA IS

A minimal microhypervisor for trusted execution environments.

NOVA is a microhypervisor designed for trusted isolation and secure shared-service execution.

It supports fully isolated virtual machines with:

  • up to 256 TB of physical memory

  • up to 128 PB of virtual address space per workload

The architecture minimizes the trusted computing base while enforcing isolation beneath the guest operating system.

This allows protections to remain intact even if workloads running above become compromised.

Three Principles

01
Minimal Trusted Base
A smaller trusted layer reduces verification complexity and minimizes attack surface.
02
Hardware-Enforced Isolation
Isolation is enforced through CPU and IOMMU protections rather than software policy alone.
03
Open and Inspectable
NOVA is released under GPLv2 and designed to be independently reviewable by the broader technical community.
01

Minimal Trusted Base

A smaller trusted layer reduces verification complexity and minimizes attack surface.

02

Hardware-Enforced Isolation

Isolation is enforced through CPU and IOMMU protections rather than software policy alone.

03

Open and Inspectable

NOVA is released under GPLv2 and designed to be independently reviewable by the broader technical community.

AMD DMA REMAPPING

DMA remapping support on AMD platforms.

The latest NOVA release introduces DMA remapping support using AMD IOMMU virtualization capabilities.

The feature is enabled as a core enforcement mechanism rather than an optional add-on.

Capability
Description
Per-VM device isolation
Prevents devices assigned to one VM from accessing neighboring workload memory.
Hardware-layer access control
Enforces memory controls directly at the hardware layer.
Per-device, per-page granularity
Restricts access at fine-grained memory boundaries.
Direct IOMMU enforcement
Blocks unauthorized DMA transactions directly through the IOMMU.
Optional fault recording
Supports DMA fault visibility for diagnostics and analysis.
Per-VM device isolation
Prevents devices assigned to one VM from accessing neighboring workload memory.
Hardware-layer access control
Enforces memory controls directly at the hardware layer.
Per-device, per-page granularity
Restricts access at fine-grained memory boundaries.
Direct IOMMU enforcement
Blocks unauthorized DMA transactions directly through the IOMMU.
Optional fault recording
Supports DMA fault visibility for diagnostics and analysis.

FORMAL VERIFICATION

Designed for provable isolation properties.

As AI infrastructure becomes more dynamic and continuously active, infrastructure trust increasingly depends on reducing implicit assumptions inside the trusted layer.

NOVA’s architecture is designed to support formal verification of the trusted computing base and minimize trusted complexity wherever possible.

The goal is not simply policy enforcement above workloads, but stronger isolation guarantees beneath them.

OPEN SOURCE

Open source and independently inspectable.

NOVA is released under the GPLv2 license.

The mechanisms responsible for enforcing infrastructure isolation should be independently reviewable, testable, and inspectable by the broader systems and security community.

Open Source Infrastructure

Trusted isolation, in the open.

NOVA is designed for teams evaluating hardware-enforced workload isolation, secure shared-service execution, and inspectable infrastructure for modern AI environments.

Open Source Infrastructure
Trusted isolation, in the open.
NOVA is designed for teams evaluating hardware-enforced workload isolation, secure shared-service execution, and inspectable infrastructure for modern AI environments.
View on GitHub Talk to the Team

What is NOVA?

NOVA is an open-source microhypervisor designed for trusted isolation and secure shared-service execution in AI infrastructure environments.

What does AMD DMA remapping support provide?

It uses hardware IOMMU functionality to restrict device memory access across isolated workloads and prevent unauthorized DMA transactions.

Is NOVA open source?

Yes. NOVA is released under the GPLv2 license and available publicly on GitHub.

Who is NOVA designed for?

Infrastructure operators, systems engineers, cloud providers, and organizations operating large-scale AI workloads that require strong workload isolation guarantees.

Why does hardware-enforced isolation matter?

Hardware-enforced isolation helps reduce dependence on software-layer trust assumptions and improves separation between workloads operating on shared infrastructure.