Help Center> > Service Overview > ECSs >GPU-accelerated ECSs

GPU-accelerated ECSs

GPU-accelerated ECSs provide outstanding floating-point computing capabilities. They are suitable for scenarios that require real-time, highly concurrent massive computing.

GPU-accelerated ECSs are classified as graphics-accelerated (G series) and computing-accelerated (P series) ECSs.
  • G series of ECSs are suitable for 3D animation rendering and CAD.
  • P series of ECSs are designed for deep learning, scientific computing, and CAE.

GPU-accelerated ECS Types

G and P series of ECSs are as follows:

Table 1 GPU-accelerated ECSs

Classification

ECS Type

GPU

Application Scenario

AZ

Remarks

G series

G5

NVIDIA V100 (GPU virtualization)

Cloud desktop, image rendering, 3D visualization, and heavy-load graphics design

CN North-Beijing 1, AZ 2

CN North-Beijing 4, AZ 1

CN East-Shanghai 2, AZ 2

CN South-Guangzhou, AZ 3

Remote login on the management console is unavailable. To log in to such an ECS, use VNC or third-party VDI.

G3

NVIDIA M60 (GPU passthrough)

CN North-Beijing 1, AZ 1

CN East-Shanghai 2, AZ 2

CN South-Guangzhou, AZ 2

Remote login on the management console is unavailable. To log in to such an ECS, use VNC or third-party VDI.

G1

NVIDIA M60 (GPU virtualization)

CN North-Beijing 1, AZ 2

CN East-Shanghai 2, AZ 1

CN South-Guangzhou, AZ 1

G1 ECSs of flavor g1.2xlarge.8 use GPU passthrough. Remote login on the management console is unavailable. To log in to such an ECS, use VNC or third-party VDI.

P series

P2v

NVIDIA V100 NVLink (GPU passthrough)

Machine learning, deep learning, reference training, scientific computing, seismic analysis, computing finance, rendering, multimedia encoding and decoding

CN North-Beijing 1, AZ 2

CN North-Beijing 4, AZ 1

CN East-Shanghai 2, AZ 2

None

P1

NVIDIA P100 (GPU passthrough)

CN North-Beijing 1, AZ 1

CN East-Shanghai 2, AZ 2

CN South-Guangzhou, AZ 2

Local NVMe SSDs are used. Pay-per-use ECSs are billed after they are stopped.

PI1

NVIDIA P4 (GPU passthrough)

CN North-Beijing 1, AZ 1

None

NOTE:
  • Remote login on the management console can be used for O&M, but it cannot be used in the production environment. Physical GPUs cannot be used if an ECS is remotely logged in through the management console.
  • If remote login on the management console is unavailable, use Windows MSTSC or a third-party desktop protocol, such as TeamViewer or VNC for login.

Images Supported by GPU-accelerated ECSs

Table 2 Image list

Classification

ECS Type

Supported Image

G series

G5

Windows Server 2012 R2 Standard 64bit

Windows Server 2016 Standard 64bit

CentOS 7.5 64bit

Ubuntu 16.04 64bit

G3

Windows Server 2012 R2 Standard 64bit

Windows Server 2008 R2 Enterprise

G1

Windows Server 2012 R2 Standard 64bit

Windows Server 2008 R2 Enterprise

P series

P2v

Windows Server 2016 Standard 64bit

Windows Server 2012 R2 Standard 64bit

Ubuntu 16.04 64bit

CentOS 7.4 64bit

EulerOS 2.2 64bit

P1

Windows Server 2012 R2 Standard 64bit

CentOS 7.3 64bit

EulerOS 2.2 64bit

Ubuntu 16.04 64bit

Debian 8.0.0 64bit

PI1

CentOS 7.3 64bit

Ubuntu 16.04 64bit

Ubuntu 14.04 64bit

Graphics-accelerated Enhancement G5

Overview

G5 ECSs use NVIDIA Tesla V100 GPUs and support DirectX, OpenGL, and Vulkan. These ECSs provide 16 GB of GPU memory and up to 4,096 x 2,160 resolution, meeting requirements from professional graphics processing.

Select your required GPU-accelerated ECS type and specifications.

Specifications
Table 3 G5 ECS specifications

Flavor

vCPUs

Memory (GB)

Maximum/Assured Bandwidth (Gbit/s)

Max. PPS (10,000)

NIC Multi-Queue

GPUs

GPU Memory (GB)

Virtualization Type

g5.8xlarge.4

32

128

25/15

200

16

V100-16Q

16

KVM

NOTE:

A g5.8xlarge.4 ECS exclusively uses a V100 GPU for professional graphics acceleration. Such an ECS can be used for heavy-load CPU reasoning.

G5 ECS Features
  • Graphics acceleration APIs
    • DirectX 12, Direct2D, DirectX Video Acceleration (DXVA)
    • OpenGL 4.5
    • Vulkan 1.0
  • CUDA* and OpenCL
  • NVIDIA V100 GPUs
  • Graphics acceleration applications
  • Heavy-load CPU reasoning
  • Application flow identical to common ECSs
  • Automatic scheduling of G5 ECSs to AZs where NVIDIA V100 GPUs are used
  • A maximum specification of 16 GB of GPU memory and 4,096 x 2,160 resolution for processing graphics and videos

Supported Common Software

G5 ECSs are used in graphics acceleration scenarios, such as video rendering, cloud desktop, and 3D visualization. If the software relies on GPU DirectX and OpenGL hardware acceleration, use G5 ECSs. G5 ECSs support the following commonly used graphics processing software:
  • AutoCAD
  • 3DS MAX
  • MAYA
  • Agisoft PhotoScan
  • ContextCapture

Notes

  • G5 ECSs do not support specifications modification.
  • G5 ECSs support the following OSs:
    • Windows Server 2012 R2 Standard 64bit
    • Windows Server 2016 Standard 64bit
    • CentOS 7.5 64bit
    • Ubuntu 16.04 64bit

Graphics-accelerated Enhancement G3

Overview

G3 ECSs are based on PCI passthrough and exclusively use GPUs for professional graphics acceleration. In addition, G3 ECSs use NVIDIA Tesla M60 GPUs and support DirectX and OpenGL with up to 16 GB of GPU memory and 4,096 x 2,160 resolution. They are ideal for professional graphics workstations.

Specifications
Table 4 G3 ECS specifications

Flavor

vCPUs

Memory (GB)

Maximum/Assured Bandwidth (Gbit/s)

Max. PPS (10,000)

NIC Multi-Queue

GPUs

GPU Memory (GB)

Virtualization Type

g3.4xlarge.4

16

64

8/2.5

50

2

1 x M60

1 x 8

KVM

g3.8xlarge.4

32

128

10/5

100

4

2 x M60

2 x 8

KVM

NOTE:

Every NVIDIA Tesla M60 card is equipped with two M60 GPUs, each of which provides 2,048 CUDA cores and 8 GB of GPU memory. M60 in G series of ECSs indicates M60 GPUs, but not M60 cards.

G3 ECS Features
  • Professional graphics acceleration APIs
  • NVIDIA M60 GPUs
  • Graphics acceleration applications
  • GPU passthrough
  • Application flow identical to common ECSs
  • Automatic scheduling of G3 ECSs to AZs where NVIDIA M60 GPUs are used
  • A maximum specification of 16 GB of GPU memory and 4,096 x 2,160 resolution for processing graphics and videos

Notes

  • G3 ECSs do not support specifications modification.
  • G3 ECSs support the following OSs:
    • Windows Server 2008 R2 Enterprise SP1 64bit
    • Windows Server 2012 R2 Standard 64bit
  • G3 ECSs do not support the remote login function provided by the public cloud platform. To remotely log in to such an ECS, use MSTSC to log in to it and install VNC on the ECS.
  • If a G3 ECS is created using a private image, install a GPU driver on the ECS after the ECS creation. For details, see Installing the GRID Driver.

Graphics-accelerated G1

Overview

G1 ECSs are based on NVIDIA GRID vGPUs and provide economical graphics acceleration. They use NVIDIA Tesla M60 GPUs and support DirectX and OpenGL. The ECSs have up to 8 GB of GPU memory and 4,096 x 2,160 resolution, and are used for applications that require high performance in graphics rendering.

Specifications

Table 5 G1 ECS specifications

Flavor

vCPUs

Memory (GB)

Maximum/Assured Bandwidth

Maximum PPS

GPUs

GPU Memory (GB)

Virtualization Type

g1.xlarge

4

8

Medium

Medium

1 x M60-1Q

1

Xen

g1.xlarge.4

4

16

Medium

Medium

1 x M60-1Q

1

Xen

g1.2xlarge

8

16

Medium

Medium

1 x M60-2Q

2

Xen

g1.2xlarge.8

8

64

Medium

Medium

Passthrough

8

Xen

g1.4xlarge

16

32

Medium

Medium

1 x M60-4Q

4

Xen

NOTE:

M60-xQ support vGPUs. x can be 1, 2, 4, or 8, indicating that M60 GPUs are virtualized to vGPUs with different specifications and models using GRID. x specifies the vGPU memory, and Q indicates that the vGPU of this type is designed to work in workstations and desktop scenarios. For more details about GRID vGPUs, see GRID VIRTUAL GPU User Guide.

G1 ECS Features
  • NVIDIA M60 GPUs
  • Graphics acceleration applications
  • GPU hardware virtualization (vGPUs) and GPU passthrough
  • Application flow identical to common ECSs
  • Automatic scheduling of G1 ECSs to AZs where NVIDIA M60 GPUs are used
  • A maximum specification of 8 GB of GPU memory and 4,096 x 2,160 resolution for processing graphics and videos
Notes
  • G1 ECSs do not support specifications modification.
  • G1 ECSs support the following OSs:
    • Windows Server 2008 R2 Enterprise SP1 64bit
    • Windows Server 2012 R2 Standard 64bit
  • g1.2xlarge.8 G1 ECSs do not support the remote login function provided by the public cloud platform. To remotely log in to such an ECS, use MSTSC to log in to it and install VNC on the ECS.

    Non-g1.2xlarge.8 G1 ECSs support the remote login function provided by the public cloud platform. For details, see Login Using VNC.

  • If a G1 ECS is created using a private image, install a GPU driver on the ECS after the ECS creation. For details, see Installing the GRID Driver.

Computing-accelerated P2v

Overview

P2v ECSs use NVIDIA Tesla V100 GPUs (16 GB GPU memory) and provide flexibility, high-performance computing, and cost-effectiveness. These ECSs use GPU NVLink for direct communication between GPUs, improving data transmission efficiency. P2v ECSs provide outstanding general computing capabilities and have strengths in AI-based deep learning, scientific computing, Computational Fluid Dynamics (CFD), computing finance, seismic analysis, molecular modeling, and genomics.

Specifications

Table 6 P2v ECS specifications

Flavor

vCPUs

Memory (GB)

Maximum/Assured Bandwidth (Gbit/s)

Maximum PPS (10,000)

NIC Multi-Queue

GPUs

GPU Connection

GPU Memory (GB)

Virtualization Type

p2v.2xlarge.8

8

64

10/4

50

4

1 x V100

N/A

1 x 16 GB

KVM

p2v.4xlarge.8

16

128

15/8

100

8

2 x V100

NVLink

2 x 16 GB

KVM

p2v.8xlarge.8

32

256

25/15

200

16

4 x V100

NVLink

4 x 16 GB

KVM

p2v.16xlarge.8

64

512

30/30

400

32

8 x V100

NVLink

8 x 16 GB

KVM

P2v ECS Features
  • Up to eight NVIDIA Tesla V100 GPUs on an ECS
  • NVIDIA CUDA parallel computing and common deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
  • 15.7 TFLOPS of single-precision computing and 7.8 TFLOPS of double-precision computing
  • NVIDIA Tensor cores with 125 TFLOPS of single- and double-precision computing for deep learning
  • Up to 30 GB/s of network bandwidth on a single ECS
  • 16 GB of HBM2 GPU memory with a bandwidth of 900 GB/s
  • Comprehensive basic capabilities

    Networks are user-defined, subnets can be divided, and network access policies can be configured as needed. Mass storage is used, elastic capacity expansion as well as backup and restoration are supported to make data more secure. Auto Scaling allows you to add or reduce the number of ECSs quickly.

  • Flexibility

    Similar to other types of ECSs, P2v ECSs can be provisioned in a few minutes.

  • Excellent supercomputing ecosystem

    The supercomputing ecosystem allows you to build up a flexible, high-performance, cost-effective computing platform. A large number of HPC applications and deep-learning frameworks can run on P2v ECSs.

Supported Common Software

P2v ECSs are used in computing acceleration scenarios, such as deep learning training, inference, scientific computing, molecular modeling, and seismic analysis. If the software requires GPU CUDA parallel computing, use P2v ECSs. P2v ECSs support the following commonly used software:
  • Common deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
  • CUDA GPU rendering supported by RedShift for Autodesk 3dsMax and V-Ray for 3ds Max
  • Agisoft PhotoScan
  • MapD
Notes
  • If a P2v ECS is created using a private image, make sure that the NVIDIA driver has been installed during the private image creation. If not, install the driver after the P2v ECS is created for computing acceleration. For details, see Installing the NVIDIA GPU Driver and CUDA Toolkit.
  • P2v ECSs do not support specifications modification.
  • Pay-per-use P2v ECSs support the following OSs:
    • Ubuntu Server 16.04 64bit
    • CentOS 7.4 64bit
    • EulerOS 2.2 64bit
    • Windows Server 2012 R2 Standard 64bit
    • Windows Server 2016 Standard 64bit

Computing-accelerated P1

Overview

P1 ECSs use NVIDIA Tesla P100 GPUs and provide flexibility, high performance, and cost-effectiveness. These ECSs support GPU Direct for direct communication between GPUs, improving data transmission efficiency. P1 ECSs provide outstanding general computing capabilities and have strengths in deep learning, graphic databases, high-performance databases, Computational Fluid Dynamics (CFD), computing finance, seismic analysis, molecular modeling, and genomics. They are designed for scientific computing.

Specifications

Table 7 P1 ECS specifications

Flavor

vCPUs

Memory (GB)

Max./Assured Bandwidth (Gbit/s)

Max. PPS (10,000)

NIC Multi-Queue

GPUs

GPU Memory (GB)

Local Disks

Virtualization Type

p1.2xlarge.8

8

64

5/1.6

35

2

1 x P100

1 x 16

1 x 800 GB NVMe

KVM

p1.4xlarge.8

16

128

8/3.2

70

4

2 x P100

2 x 16

2 x 800 GB NVMe

KVM

p1.8xlarge.8

32

256

10/6.5

140

8

4 x P100

4 x 16

4 x 800 GB NVMe

KVM

P1 ECS Features
  • Up to four NVIDIA Tesla P100 GPUs on a P1 ECS (If eight P100 GPUs are required on an instance, use BMS.)
  • GPU hardware passthrough
  • 9.3 TFLOPS of single-precision computing and 4.7 TFLOPS of double-precision computing
  • Maximum network bandwidth of 10 Gbit/s
  • 16 GB of HBM2 GPU memory with a bandwidth of 732 GB/s
  • 800 GB NVMe SSDs for temporary local storage
  • Comprehensive basic capabilities

    Networks are user-defined, subnets can be divided, and network access policies can be configured as needed. Mass storage is used, elastic capacity expansion as well as backup and restoration are supported to make data more secure. Auto Scaling allows you to add or reduce the number of ECSs quickly.

  • Flexibility

    Similar to other types of ECSs, P1 ECSs can be provisioned in a few minutes. You can configure specifications as needed. P1 ECSs with two, four, and eight GPUs will be supported later.

  • Excellent supercomputing ecosystem

    The supercomputing ecosystem allows you to build up a flexible, high-performance, cost-effective computing platform. A large number of HPC applications and deep-learning frameworks can run on P1 ECSs.

Supported Common Software

P1 ECSs are used in computing acceleration scenarios, such as deep learning training, inference, scientific computing, molecular modeling, and seismic analysis. If the software requires GPU CUDA parallel computing, use P1 ECSs. P1 ECSs support the following commonly used software:

  • Deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
  • RedShift for Autodesk 3dsMax, V-Ray for 3ds Max
  • Agisoft PhotoScan
  • MapD
Notes
  • It is recommended that the system disk of a P1 ECS be greater than 40 GB.
  • P1 ECSs have local NVMe SSDs attached, which are still billed after the ECSs are stopped. To stop billing such an ECS, delete it.
  • The local NVMe SSDs attached to P1 ECSs are dedicated for services with strict requirements on storage I/O performance, such as deep learning training and HPC. Local disks are attached to the ECSs of specified flavors and cannot be separately bought. In addition, you are not allowed to detach a local disk and then attach it to another ECS.
    NOTE:

    Data may be lost on the local NVMe SSDs attached to P1 ECSs due to a fault, for example, due to a disk or host fault. Therefore, you are suggested to store only temporary data in local NVMe SSDs. If you store important data in such a disk, securely back up the data.

  • If a P1 ECS is created using a private image, make sure that the NVIDIA driver has been installed during the private image creation. If not, install the driver after the P1 ECS is created for computing acceleration. For details, see Installing the NVIDIA GPU Driver and CUDA Toolkit.
    NOTE:

    For instructions about how to create a private image, see Image Management Service User Guide.

  • P1 ECSs do not support specifications modification.
  • P1 ECSs do not support automatic recovery.
  • P1 ECSs support the following OSs:
    • Debian 8.0 64bit
    • Ubuntu Server 16.04 64bit
    • CentOS 7.3 64bit
    • EulerOS 2.2 64bit
    • Windows Server 2012 R2 Standard 64bit
  • After you delete a P1 ECS, the data stored in local NVMe SSDs is automatically cleared.

Inference-accelerated PI1

Overview

PI1 ECSs use NVIDIA Tesla P4 GPUs dedicated for real-time AI inference. Working with P4 INT8 calculators, PI1 ECSs have shortened the inference latency by 15 times. Working with hardware decoding engines, PI1 ECSs concurrently support real-time 35-channel HD video transcoding and inference.

Specifications

Table 8 PI1 ECS specifications

Flavor

vCPUs

Memory (GB)

Max./Assured Bandwidth (Gbit/s)

Maximum PPS (10,000)

NIC Multi-Queue

GPUs

GPU Memory (GB)

Local Disks

Virtualization Type

pi1.2xlarge.4

8

32

5/1.6

40

2

1 x P4

1 x 8 GB

N/A

KVM

pi1.4xlarge.4

16

64

8/3.2

70

4

2 x P4

2 x 8 GB

N/A

KVM

pi1.8xlarge.4

32

128

10/6.5

140

8

4 x P4

4 x 8 GB

N/A

KVM

PI1 ECS Features
  • Up to four NVIDIA Tesla P4 GPUs on an ECS
  • GPU hardware passthrough
  • Up to 5.5 TFLOPS of single-precision computing on a single GPU
  • Up to 22 TOPS of INT8 computing on a single GPU
  • 8 GB of ECC GPU memory with a bandwidth of 192 GB/s on a single GPU
  • Hardware video encoding and decoding engines embedded in GPUs for concurrent real-time 35-channel HD video transcoding and inference

Supported Common Software

PI1 ECSs are used in GPU-based inference computing scenarios, such as image recognition, voice recognition, and natural language processing.

PI1 ECSs support the following commonly used software:

  • Deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
Notes
  • The basic resources, including vCPUs, memory, and image of a pay-per-use PI1 ECS of flavor pi1.2xlarge.4, pi1.4xlarge.4, or pi1.8xlarge.4 are not billed after the ECS is stopped, but the system disk of the ECS is still being billed according to the disk capacity. If the ECS is bound with other resources, such as EVS disks, EIPs, and bandwidths, these resources are billed using their own billing mode (yearly/monthly or pay-per-use). For details, see Product Pricing Details.

    The resources of a pay-per-use PI1 ECS of flavor pi1.2xlarge.4, pi1.4xlarge.4, or pi1.8xlarge.4 are released after the ECS is stopped. If the underlying resources are insufficient when the ECS is started, starting the ECS may fail. If you want to use such an ECS for a long period of time, you are advised to keep the ECS running or select the yearly/monthly payment.

  • If a PI1 ECS is created using a private image, make sure that the GPU driver has been installed during the private image creation. If not, install the driver after the PI1 ECS is created for inference acceleration. For details, see Installing the NVIDIA GPU Driver and CUDA Toolkit.
  • PI1 ECSs support the following OSs:
    • Ubuntu Server 14.04 64bit
    • Ubuntu Server 16.04 64bit
    • CentOS 7.3 64bit
  • PI1 ECSs do not support specifications modification.
  • PI1 ECSs support automatic recovery when the hosts accommodating such ECSs become faulty.