BlueRock's NOVA Microhypervisor Adds AMD DMA Isolation for AI
New NOVA release blocks DMA attacks with AMD IOMMU, securing shared AI workloads
BlueRock has released the latest open-source version of its NOVA Microhypervisor, now with DMA remapping support for AMD platforms featuring IOMMU hardware virtualization. This capability, enabled by default, extends hardware-level isolation across virtual machines, devices, and memory in shared execution environments. The microhypervisor combines microkernel and hypervisor functions in a small trusted computing base, using a capability-based authorization model. It supports ARMv8-A and x86_64 processors, including Intel VT-x and AMD-V. The new AMD IOMMU integration can prevent hardware devices assigned to one VM from accessing memory of neighboring workloads, enforce per-device and per-page memory access controls, and abort unauthorized DMA transactions.
The NOVA microhypervisor also supports VMs with up to 256TB of physical memory and 128PB of virtual address space per workload, using deep 5-level radix trees for page tables that are completely lockless. Protection Domains can be isolated to specific CPU cores for predictable AI workload execution, and CPU caches can be partitioned for quality-of-service optimization. On x86, optional Control-Flow Enforcement Technology (Intel CET) is available but omitted by default due to overhead. The project remains experimental and is licensed under GPL v2, with code available on GitHub. BlueRock positions this as essential for continuously running AI production infrastructure requiring isolation, predictability, and reduced trusted complexity.
- AMD IOMMU DMA remapping prevents compromised device drivers from reading or writing memory of neighboring VMs.
- Supports VMs with up to 256TB physical memory and 128PB virtual address space per workload.
- Lockless page tables enable scalable concurrent updates to disjoint memory regions for AI workloads.
Why It Matters
As AI infrastructure goes production, this microhypervisor isolates workloads at the hardware level, mitigating chipset-based attacks.