User‑Mode vs Kernel‑Mode GPU Virtualization: Architecture, Benefits, and Limits

User‑Mode Virtualization

In user‑mode virtualization the application uses standard CUDA, OpenGL, and similar APIs; a user‑space library intercepts and forwards these calls, optionally using RPC to enable remote GPU invocation.

Open standard interfaces provide openness and stability.

Running in user space avoids complex kernel code and associated security issues.

Minimal intrusion on the host environment makes deployment safest.

Supports unified memory, allowing the GPU to borrow host RAM.

Drawback: the number of API functions is orders of magnitude larger than in kernel‑mode interfaces.

Kernel‑Mode Virtualization

Kernel‑mode solutions intercept kernel‑level interfaces such as ioctl, mmap, read, and write to virtualize the GPU.

Higher flexibility; does not depend on specific GPU hardware and works on both data‑center and consumer GPUs.

Provides good isolation while sharing GPUs.

Implemented in container environments, reducing development effort compared with user‑mode solutions.

Drawbacks: requires inserting kernel modules, increasing system intrusion and security risk; private‑cloud customers have heterogeneous kernel versions, raising compatibility and security concerns; reverse‑engineering proprietary interfaces creates legal uncertainty and reduces sustainability; currently limited to container usage; cannot use unified memory.

Cross‑Process Interface Parsing

After API interception, two parsing models exist:

Same‑process space parsing : works only within a single process; it cannot support KVM virtualization across OS kernels or remote GPU calls across physical nodes.

Different‑process space parsing : because the application and GPU driver run under different OS instances, each of the tens of thousands of CUDA interfaces must be simulated, making the approach complex.

Overall, user‑mode and kernel‑mode GPU virtualization each have distinct strengths and limitations that affect their suitability for remote GPU invocation and resource pooling.