GPU Requirements

Either a CUDA (12 or later) or an OpenCL driver (1.2 or later) for the GPU is necessary to use Zivid.

• For Windows, the driver is usually installed together with the graphics driver: NVIDIA, Intel.

• For Linux, install the GPU driver for your backend: see About GPU for installing the NVIDIA driver (CUDA backend) or the OpenCL driver.

Follow the instructions on how to check whether you have OpenCL. For the CUDA backend, check that nvidia-smi reports an NVIDIA driver that supports CUDA 12 or later.

Dedicated GPU

A dedicated GPU gives the best performance with Zivid. It is also the best choice if the GPU will be used for more than Zivid’s computations. A medium to high-end NVIDIA GPU with at least 3 GB of memory is required for optimal performance. This is the preferred solution for Zivid.

Requirements:

• NVIDIA GPUs must meet a minimum compute capability, which differs between the two compute backends:

  • CUDA backend: compute capability 6.1 or higher. This covers the GTX 10-series and everything newer, so cards from 2016 onwards. The exception lists the compute capabilities the installed package was built for:

    Detected compute capability X.Y, which is not supported. Supported versions: (6.1, 7.0, 7.5, 8.0, 8.6)

    Setting AllowUnsupported does not make the CUDA backend work below this minimum. The packages contain no code for older architectures, so the capture fails later instead. Use a newer GPU, or the OpenCL version of the SDK.

  • OpenCL backend: no practical limit. The requirement is OpenCL 1.2 support, which every NVIDIA card new enough to have an OpenCL driver provides. A compute capability floor of 3.5 exists in the SDK, but no driver that meets the OpenCL requirement falls below it, so in practice the OpenCL version covers the full range of NVIDIA cards:

    The compute capability X.Y of the Nvidia compute device is lower than the minimum required compute capability 3.5. Update to a supported compute device by consulting <link>.

  A GPU below 6.1 therefore works with the OpenCL backend but not with the CUDA backend. Be aware that the OpenCL version of the SDK rejects every NVIDIA GPU regardless of compute capability until you opt in to unsupported compute devices in the Config.yml file. The 3.5 minimum applies only once that opt-in is in place.

  On Jetson the CUDA package is built for the compute capability of the Jetson hardware itself, so this minimum does not apply, and the OpenCL version is not available. See Install Zivid on Jetson Linux.

  An overview of NVIDIA GPUs and their corresponding compute capability can be found in https://developer.nvidia.com/cuda-gpus.

Recommendations:

• NVIDIA GeForce GTX 1060 or newer for desktops

• NVIDIA GeForce MX150 or newer for laptops

  Caution

  Support for Ubuntu 18.04 is removed since SDK 2.10. To see Ubuntu 18.04 instructions for an earlier SDK version, change the Knowledge Base version in the top left corner.

CPU with integrated GPU

This is the most cost-efficient solution for Zivid. The benefit of an integrated GPU is that the data transfer from GPU to CPU is fast. It is also beneficial for applications that require low power consumption, e.g., battery-powered mobile platforms. A high-end integrated GPU with at least 3 GB of memory available is required for optimal performance.

Recommendations:

• Intel i7 with HD630 or better

  Caution

  Support for Ubuntu 18.04 is removed since SDK 2.10. To see Ubuntu 18.04 instructions for an earlier SDK version, change the Knowledge Base version in the top left corner.

Tip

Use CPU with integrated GPU for applications that require low power consumption, e.g., battery powered mobile platforms with cameras.

Version History

SDK	Changes
2.18.0	Added support for the CUDA compute backend for NVIDIA GPUs.
2.11.0	AMD is no longer recommended. Official support has been removed in SDK, see Enable Unsupported Compute Device for Zivid Software.
2.10.0	Support for Ubuntu 18.04 removed.