Do not equate CUDA capability with a graphics-rendering requirement

Source facts

CUDA uses a GPU’s general-purpose compute resources rather than its graphics rasterization pipeline. Microsoft documents CUDA on both compute-only and graphics-capable NVIDIA GPU VM families. It is a software stack containing drivers, compiler/runtime components, libraries, and tools, not just a driver. The validated OS, driver, and Toolkit combinations depend on the Toolkit release and GPU. Microsoft Learn.

Applicability

Apply this distinction when a request says only that an application needs a GPU or CUDA. Ask whether the application performs numerical computation, renders interactive graphics, or needs both. Do not treat a framework name as a complete hardware specification.

DSE recommendation

Separate the workload’s compute requirements from its display and rendering requirements. Have the application owner identify the actual operations and supported software combination before comparing VM sizes. Review the size specifications and the current compatibility documentation for that combination. Keep a deliberate record of any graphics requirement so it is neither omitted from a visualization deployment nor assumed for a compute-only job.

Verification

On the proposed configuration, run a representative application task that exercises the required acceleration path. Include an actual rendering task if graphics are part of the requirement, and keep the runtime and driver versions with the result. Compare application behavior with the agreed acceptance criteria rather than using successful driver installation as the sole approval.

Official references

Microsoft Learn: NVIDIA CUDA on Azure overview. Source reviewed September 9, 2026.