Graphics accelerators (GPUs)
Compute Cloud provides graphics accelerators (GPUs) for different VM configurations. GPUs outperform CPUs for certain types of data and can be used for complex computing. For even more performance and convenience, you can use automatic allocation of resources in Yandex DataSphere.
The following GPUs are available in Compute Cloud:
- NVIDIA® Tesla® V100 with 32 GB HBM2 (High Bandwidth Memory).
- NVIDIA® Ampere® A100 with 80 GB HBM2.
- NVIDIA® Tesla® T4 with 16 GB GDDR6.
Warning
GPUs run in TCC mode, which does not use the operating system's graphics drivers.
By default, the cloud has a zero quota for creating VMs with GPUs. You can request a quota increase in the management console. To do this, you need the quota-manager.requestOperator role or higher.
You cannot create VMs with GPUs in the ru-central1-c availability zone. For more information, see Discontinuation of the ru-central1-c availability zone.
Graphics accelerators (GPUs)
Graphics accelerators are suitable for machine learning (ML), artificial intelligence (AI), and 3D rendering tasks.
You can manage GPUs and RAM directly from your VM.
NVIDIA® Tesla® V100
The NVIDIA® Tesla® V100 graphics card contains 5120 CUDA® cores for high-performance computing (HPC), and 640 Tensor cores for deep learning (DL) tasks.
NVIDIA® Ampere® A100
The NVIDIA® A100 GPU based on the Ampere® microarchitecture uses third-generation Tensor Cores and offers 80 GB HBM2 memory with up to 2 TB/s bandwidth.
NVIDIA® Tesla® T4
NVIDIA® Tesla® T4 based on the Turing™ architecture uses Turing tensor cores and RT cores and offers 16 GB of GDDR6 memory with 300 GB/s bandwidth.
VM configurations
The computing resources may have the following configurations:
-
Intel Broadwell with NVIDIA® Tesla® V100 (
gpu-standard-v1):Number of GPUs VRAM, GB Number of vCPUs RAM, GB 1 32 8 96 2 64 16 192 4 128 32 384 -
Intel Cascade Lake with NVIDIA® Tesla® V100 (
gpu-standard-v2):Number of GPUs VRAM, GB Number of vCPUs RAM, GB 1 32 8 48 2 64 16 96 4 128 32 192 8 256 64 384 -
AMD EPYC™ with NVIDIA® Ampere® A100 (
gpu-standard-v3):Number of GPUs VRAM, GB Number of vCPUs RAM, GB 1 80 28 119 2 160 56 238 4 320 112 476 8 640 224 952 -
AMD EPYC™ 9474F with Gen2 (
gpu-standard-v3i):Number of GPUs VRAM, GB Number of vCPUs RAM, GB 8 640 180 1,440 -
Intel Ice Lake with NVIDIA® Tesla® T4 (
standard-v3-t4):Number of GPUs VRAM, GB Number of vCPUs RAM, GB 1 16 4 16 1 16 8 32 1 16 16 64 1 16 32 128
VM GPUs are provided in full. For example, if a configuration has four GPUs specified, your VM will have four full-featured GPU devices.
You can create VMs on Intel Broadwell with NVIDIA® Tesla® V100, Intel Cascade Lake with NVIDIA® Tesla® V100, AMD EPYC™ with NVIDIA® Ampere® A100 and Intel Ice Lake with NVIDIA® Tesla® T4 in the ru-central1-a and ru-central1-b availability zones.
For more information about organizational and technical limitations for VMs, see Quotas and limits.
For information about the cost of VMs with GPUs, see VM computing resources.
OS images
For VMs with GPUs, you can use the following special images of operating systems with NVIDIA drivers:
- Intel Broadwell with NVIDIA® Tesla® V100 and Intel Cascade Lake with NVIDIA® Tesla® V100
-
- Ubuntu 16.04 LTS GPU (
ubuntu-1604-lts-gpu) - Ubuntu 20.04 LTS GPU (
ubuntu-2004-lts-gpu)
- Ubuntu 16.04 LTS GPU (
- Intel Ice Lake with NVIDIA® Tesla® T4
-
- Ubuntu 20.04 LTS GPU (
ubuntu-2004-lts-gpu)
- Ubuntu 20.04 LTS GPU (
- AMD EPYC™ with NVIDIA® Ampere® A100
-
- Ubuntu 20.04 LTS GPU A100 (
ubuntu-2004-lts-a100) - Ubuntu 18.04 LTS GPU A100 (
ubuntu-1804-lts-a100)
- Ubuntu 20.04 LTS GPU A100 (
We recommend using a standard image from Yandex Cloud. You can also install the drivers on another standard image yourself or create a custom image with pre-installed drivers.
GPU clusters
You can group several VMs into a cluster. This will allow you to accelerate distributed training tasks that require higher computing capacity than individual VMs can provide. Make sure the cluster is created in the same availability zone as its VMs. The cluster VMs are interconnected through InfiniBand, a secure high-speed network.
You can add VMs from different folders, networks, and subnets to your cluster. For the cluster VMs to interact properly, we recommend using a security group that allows unlimited traffic within the group. The default security group meets this requirement. If you edited the default security group, add a group with unlimited internal traffic.
Maximum cluster size for AMD EPYC™ 9474F with Gen2 is 20 VMs with 8 GPU, 80 GB VRAM, 180 vCPU, 1,440 GB RAM configuration. The actual maximum cluster size is limited by the technical availability of the resources.