40. GPU computations

40.1. Introduction

AGX Dynamics has no hard requirement for a GPU or for specific GPU drivers to be present. However, some AGX features can use CUDA for GPU computations.

When such features are used, additional AGX libraries are loaded dynamically. This means that the regular AGX libraries do not require CUDA or specific GPU drivers to be installed. If a supported graphics card is not available, or if the required drivers are missing, dynamic loading will fail gracefully, and the affected features will not be available or enabled.

GPU computations and dynamic library loading are used by:

  • AGX Sensor - for lidar simulation.

  • AGX Granular - to accelerate the iterative contact solver.

Because multiple libraries provide GPU support, they link against a shared CUDA runtime to avoid embedding multiple static runtimes. This runtime, cudart, is redistributed with the AGX installer under the license grant in the NVIDIA CUDA End User License Agreement.

40.2. Hardware requirements

CUDA Toolkit 13 is used when building GPU components for AGX. On Windows, version 13.2 is used. On Ubuntu version 13.1 is used. The CUDA 13.x toolkit family requires NVIDIA GPU driver version of 580 or later, and hardware with Compute Capability 7.5 or higher.

For a detailed list of supported GPUs, see CUDA GPU Compute Capability.

AGX 2.42 and earlier used CUDA Toolkit 12.x, which required driver version 525 or later and hardware with at least Compute Capability 5.0.

40.3. Fractional GPUs

When running computations in a cloud environment, virtualized GPUs may be used. The underlying virtualization technology can affect the running program, either by introducing additional latency when GPU access is time-sliced, or by physically partitioning the GPU.

These types of GPUs are supported by AGX Sensor and can be used for lidar simulation. In most cases, they will have no noticeable impact on simulation performance.

Virtualized shared GPUs are not recommended for granular simulation, where performance depends heavily on memory bandwidth and the amount of available compute capacity.