BlueRock’s NOVA microhypervisor adds AMD DMA isolation to harden shared AI infrastructure

BlueRock, a microhypervisor developer, announced a new release of its NOVA platform that adds AMD-specific direct memory access (DMA) remapping support for systems using AMD IOMMU hardware. DMA is how devices read and write data in memory without involving the CPU for every transfer. DMA remapping, as BlueRock describes it, is intended to block unwanted cross-workload or cross-virtual-machine access when multiple AI services share underlying infrastructure.

The company said the NOVA microhypervisor is designed for “secure-shared execution,” an approach meant to combine isolation with efficiency as AI workloads move from experimental runs to continuously active production services. In that setting, BlueRock argues that execution coordination and system architecture become as important as raw compute performance, because operators face rising costs and strain from running inference and related services at scale.

BlueRock’s announcement ties the update to hardware-level controls, saying the new AMD DMA remapping capability strengthens isolation across workloads, devices, and memory in shared execution environments. It also characterizes the release as part of an effort to support fully isolated virtual machines with up to 256TB of physical memory and 128 petabytes of virtual address space per workload, a scale target aimed at large AI infrastructure deployments.

In the same post, BlueRock says NOVA was built to deliver trusted isolation and secure shared-service execution at large AI infrastructure scale. That includes the ability to keep each workload protected while still allowing multiple services to share underlying platforms, an architecture that would otherwise require either heavy overprovisioning or complex software controls to meet security and performance needs.

For AMD platforms, BlueRock’s emphasis is on IOMMU hardware virtualization capabilities, which are commonly used to restrict what devices can access. By adding advanced DMA remapping support, the company is effectively pushing isolation closer to the hardware boundary, where security teams typically prefer controls to be enforced rather than relying entirely on software policies.

The practical business context is that many operators are trying to reduce the cost of running inference and other AI workloads. With AI moving toward more agentic behavior, continuous activity, and distributed systems, BlueRock frames the problem as one where concurrency and execution complexity rise faster than the economics of simply scaling compute by adding more dedicated servers.

BlueRock did not provide a list of specific customers adopting NOVA, nor did it disclose any performance benchmarks, adoption metrics, or formal validation results tied to the AMD DMA remapping feature in the announcement. It also did not specify whether the update is limited to certain CPU generations, OEM systems, or IOMMU configurations, leaving implementation details to developers integrating the release.

For AMD watchers and data-center infrastructure teams, the next question will be whether the added hardware isolation features translate into easier deployment for security-sensitive AI workloads, and whether major platform vendors or cloud operators incorporate NOVA in ways that make AMD-based secure shared execution more mainstream. Watch for release documentation that clarifies supported hardware matrices and for independent reporting on real-world operational impacts.