Newly Granted Patent: Device Virtualization for Scalable I/O Management
I’m glad to share that I’ve been granted a new US patent, US 12,650,940 B2, “Device Virtualization Techniques”, alongside co-inventors Ryan Holmqvist, Kapil Karkra, Orden Smith, Nicholas Adams, and Vinay Raghav, assigned to Intel Corporation.
The Problem
Modern systems increasingly need to manage large, dynamic pools of I/O devices behind a stable software-facing identity. NVMe storage is one example, but the same challenge shows up for GPUs and AI accelerators: large clusters of them get added, removed, reassigned, and hot-plugged as workloads scale, while drivers and orchestration software generally expect well-known, stable device identities. Solving that mismatch well requires sitting squarely at the hardware/software interface, not purely in silicon and not purely in software.
The Idea, in Brief
The patent describes a mechanism for a physical I/O device to be locked behind a virtual identity. Reads that would normally return the physical device’s identifying information are intercepted and substituted with values associated with the virtual device. The device can later be unlocked, exposing the real physical identity to a native driver, in response to a request to a predetermined address. It’s a lightweight, general building block for letting system software manage the appearance of I/O devices independently from their physical reality, without requiring a full virtualization stack to intervene on every transaction. Because the mechanism only cares about device identity and access, it applies just as well to a rack of AI accelerators as it does to a bank of NVMe drives.
Why This Matters as a System Architect
This isn’t the first time I’ve worked this class of problem. I’m the architect of record for Intel VMD (Volume Management Device), the hardware underpinning Intel VROC (Virtual RAID on CPU). VMD is, at its core, an I/O aggregation technology. It lets system software manage large numbers of NVMe drives far more easily, and it enables hot-plug and software RAID even in the presence of I/O virtualization.
This new patent tackles a kindred scalability and management challenge, but through a different, more general mechanism, and with an eye toward the kind of large, fluid device pools that AI infrastructure now demands. It’s not a replacement or extension of VMD, just a related idea from the same design space. I’m sharing it here as a patent, on its own merits, without speculating on what Intel may or may not do with it.
A First-Principles Problem
What I find most representative of how I work as an architect is the way this idea was derived, not just the claims it produced. It started from taking VMD apart down to its fundamental function and asking what actually makes it work at the hardware/software boundary. Underneath all the plumbing, the core hook is deceptively simple: control what identity a device presents to software, and you control how software organizes, pools, and manages that device, without touching the device itself or the driver stack built around it. Once that hook is isolated, the rest follows from minimalism. The mechanism in this patent is a small, cheap trick implemented largely in existing interfaces, no new datapaths, no significant hardware cost, just a deliberate substitution at the right place in the read path to fool software into seeing what the system wants it to see. That’s the trade I look for in every design: a small, well-placed mechanism with disproportionate leverage at the system level.
This is one entry in a broader body of work. The full, current list of my granted patents and pending applications is on the About page.