STRATUM GPU & Guest Tools Overview
The normal STRATUM GPU path is simple: attach STRATUM GPU, install STRATUM Guest Tools, then run ordinary CUDA applications.
Normal guest workflow
The normal STRATUM GPU path is simple: attach STRATUM GPU, install STRATUM Guest Tools, then run ordinary CUDA applications. The physical GPU can be far away—on another host, at another site, or anywhere else reachable through the STRATUM Continuum.
The guest does not require the NVIDIA kernel GPU driver for the STRATUM virtual GPU path. Normal customer setup is zero-configuration after Guest Tools installation: it should work out of the box. Manual LD_LIBRARY_PATH overrides, native-port variables, and mock-provider settings are diagnostic or development tools, not part of the supported customer path.
Interface tour
Attach STRATUM GPU
Edit the VM and enable STRATUM GPU in the AI Accelerator Fabric section. Select the GPU class/count required by the workload, then save the VM configuration.
GPU Splicing: 2–16 logical GPUs. Real CUDA.
Soft-MIG is a first-class STRATUM Accelerator Fabric feature, not a simulated GPU backend. Divide one supported physical GPU into up to 16 independently allocatable logical resources backed by the parent GPU's real CUDA compute engine. STRATUM controls lease memory in software, time-slices compute across contexts, and presents each VM with its own STRATUM GPU.
Explore GPU Splicing (Soft-MIG), commands and NVIDIA MPS comparison →
The outcome you should see
A VM can be configured with STRATUM GPU, then completed through the Guest Tools installation and validation pages without application-specific STRATUM configuration.