Unified Virtual Memory (UVM)#

Unified Virtual Memory (UVM) gives the guest a single virtual address space visible to both CPU and GPU. Host code and device code can both touch allocations in that space. The application does not need to issue an explicit copy for every transfer. This simplifies some CUDA® programs.

For behavior and enablement steps, refer to the Unified Virtual Memory documentation.

Unified Virtual Memory (UVM)

UVM Known Issues and Limitations#

  • Unified Virtual Memory (UVM) is restricted to 1:1 time-sliced and MIG vGPU™ for Compute profiles that allocate the entire framebuffer of a compatible physical GPU or GPU Instance. Fractional time-sliced vGPUs do not support UVM.

  • UVM is only supported on Linux Guest OS distros. Windows Guest OS is not supported.

  • Enabling UVM disables vGPU migration for the VM, which can limit your options in environments that rely on live migration.

  • On NVIDIA® A100 (Ampere) GPUs, you can enable UVM on a multi-vGPU VM in SR-IOV Heavy mode. Doing so disables P2P over NVLink™ and NVSwitch™ between the vGPUs assigned to that VM. Refer to Peer-to-Peer (P2P) CUDA Transfers for details. This restriction does not apply to NVIDIA Hopper™ (H100/H200) or newer architectures.

  • UVM is disabled by default. You must enable it for each vGPU that requires it. To do so, set a specific vGPU plugin parameter for the VM.

Note

When deploying NVIDIA NIM™, if UVM is enabled and an optimized engine is available, the model runs on the TensorRT™-LLM (TRT-LLM) backend. Otherwise, it typically runs on the vLLM backend.

Hypervisor Platform Support for UVM#

Unified Virtual Memory (UVM) is disabled by default. If used, you must enable unified memory individually for each vGPU for Compute VM that requires it by setting a vGPU plugin parameter. How to enable UVM for a vGPU VM depends on the hypervisor that you are using.

vGPU Support for UVM#

UVM is supported on 1:1 MIG-backed and time-sliced vGPUs. These vGPUs have the entire framebuffer of a MIG GPU Instance or physical GPU assigned to a single vGPU.

For the architecture-specific board and vGPU matrix, refer to UVM Board Support.