Graphics accelerators (GPUs)

Yandex Compute Cloud provides graphics accelerators (GPUs) as part of graphics cards. GPUs outperform vCPUs in processing certain types of data and can be used for complex computing.

Compute Cloud uses NVIDIA® Tesla® V100 GPUs with 32 GB HBM2.

The NVIDIA® Tesla® V100 graphics card contains 5120 CUDA® cores that perform high performance computing (HPC), and 640 Tensor cores for deep learning (DL) tasks.

Graphics accelerators are also suitable for machine learning (ML), artificial intelligence (AI), and 3D rendering tasks.

You can control a GPU and RAM directly from your VM.

VM configurations

Available configurations of computing resources:

• Intel Broadwell with NVIDIA® Tesla® V100 (gpu-standard-v1):

  Number of GPUs	Number of vCPUs	RAM, GB
  1	8	96
  2	16	192
  4	32	384

• Intel Cascade Lake with NVIDIA® Tesla® V100 (gpu-standard-v2):

  Number of GPUs	Number of vCPUs	RAM, GB
  1	8	48
  2	16	96
  4	32	192
  8	64	384

For more information about VM organizational and technical limits, see Quotas and limits.

OS images

For VMs with GPUs, special images of Windows (2016 Datacenter GPU, windows-2016-gvlk-gpu) and Ubuntu (16.04 lts GPU, ubuntu-1604-lts-gpu) are available with NVIDIA drivers installed. To use other images, install the necessary drivers on your own.

Virtual graphics accelerators (vGPUs)

Yandex Compute Cloud lets you virtualize graphics accelerators (GPUs). Virtual GPUs are created based on NVIDIA® vGPU technology.

NVIDIA® vGPU software lets you use cards with GPUs for both graphics and computing tasks on vGPUs. This requires the appropriate licenses.

To use vGPU technology, you need:

• A VM running on the platform vgpu-standard-v1 with one of the following images:
  • Ubuntu 18.04 lts vGPU.
  • Windows Server 2019 Datacenter vGPU.
• License to use NVIDIA® vGPU technology.
• NVIDIA® vGPU Software License Server.

To work with the license, you can:

• Use your current license server.
  The current license server must be available over the network from VMs with vGPUs.
• Create a VM with the NVIDIA® vGPU Software License Server in Yandex.Cloud.
  For information about how to install and configure the license server, see the NVIDIA documentation.

Configurations of VMs with vGPUs

The following configuration is available for VMs with vGPUs:

• Intel Broadwell with NVIDIA® vGPU Tesla® V100 8G (vgpu-standard-v1):

  Number of vGPUs	Number of vCPUs	RAM, GB	GPU RAM, GB
  1	4	12	8

See also

• Create a VM with a GPU.
• Learn how to add a GPU to an existing VM.
• Learn how to change the number of GPUs.