Help Center/ Elastic Cloud Server/ User Guide (ME-Abu Dhabi Region)/ Service Overview/ ECS Types and Specifications/ x86 ECS Specifications/ GPU-accelerated ECSs
Updated on 2025-12-10 GMT+08:00
View PDF
Share

GPU-accelerated ECSs

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

GPU-accelerated ECSs

Helpful links:
Table 1 GPU-accelerated ECSs

Type

Series

GPU

CUDA Cores per GPU

Single-GPU Performance

Application

Computing-accelerated

P2s

NVIDIA V100

5,120
  • 14 TFLOPS of single-precision floating-point computing
  • 7 TFLOPS of double-precision floating-point computing
  • 112 TFLOPS Tensor Cores for deep learning acceleration

AI deep learning training, scientific computing, computational fluid dynamics, computational finance, seismic analysis, molecular modeling, and genomics.

Inference-accelerated

Pi2

NVIDIA T4 (GPU passthrough)

2,560
  • 8.1 TFLOPS of single-precision floating-point computing
  • 130 INT8 TOPS
  • 260 INT4 TOPS

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

Inference-accelerated

Pi2nl

NVIDIA P4 (GPU passthrough)

2,560
  • 8.1 TFLOPS of single-precision floating-point computing
  • 130 INT8 TOPS
  • 260 INT4 TOPS

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

Images Supported by GPU-accelerated ECSs

Table 2 Images supported by GPU-accelerated ECSs

Type

Series

Supported Image

Computing-accelerated

P3
  • CentOS 8.2 64bit
  • CentOS 8.1 64bit
  • CentOS 8.0 64bit
  • CentOS 7.9 64bit
  • CentOS 7.8 64bit
  • CentOS 7.7 64bit
  • CentOS 7.6 64bit
  • Ubuntu 20.04 server 64bit
  • Ubuntu 18.04 server 64bit

Computing-accelerated

P2s
  • Windows Server 2016 Standard 64bit

Inference-accelerated

Pi2
  • CentOS 7.5 64bit
  • Windows Server 2019 Standard 64bit
  • Windows Server 2016 Standard 64bit

Inference-accelerated

Pi2nl
  • CentOS 7.5 64bit
  • Ubuntu 16.04 server 64bit
  • Windows Server 2016 Standard 64bit

Computing-accelerated P3

Overview

P3 ECSs use NVIDIA A100 GPUs and provide flexibility and ultra-high-performance computing. P3 ECSs have strengths in AI-based deep learning, scientific computing, Computational Fluid Dynamics (CFD), computing finance, seismic analysis, molecular modeling, and genomics. Theoretically, the FP32 is 19.5 TFLOPS, and the TF32 tensor core is 156 TFLOPS | 312 TFLOPS (sparsity enabled).

Specifications

Table 3 P3 ECS specifications

Flavor

vCPUs

Memory

(GiB)

Max./Assured Network Bandwidth (Gbit/s)

Max. Network PPS

(10,000)

Max. NIC Queues

Max. NICs

GPUs

GPU Memory

(GiB)

Virtualization

p3.2xlarge.8

8

64

10/4

100

4

4

1 × NVIDIA A100 80 GB

80

KVM

p3.4xlarge.8

16

128

15/8

200

8

8

2 × NVIDIA A100 80 GB

160

KVM

p3.8xlarge.8

32

256

25/15

350

16

8

4 × NVIDIA A100 80 GB

320

KVM

p3.16xlarge.8

64

512

36/30

700

32

8

8 × NVIDIA A100 80 GB

640

KVM

P3 ECS Features

  • CPU: 2nd Generation Intel® Xeon® Scalable 6248R processors and 3.0 GHz of basic frequency
  • Up to eight NVIDIA A100 GPUs on an ECS
  • NVIDIA CUDA parallel computing and common deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
  • 19.5 TFLOPS of single-precision computing and 9.7 TFLOPS of double-precision computing on a single GPU
  • NVIDIA Tensor cores with 156 TFLOPS of single- and double-precision computing for deep learning
  • Up to 40 Gbit/s of network bandwidth on a single ECS
  • 80 GB HBM2 GPU memory per graphics card, with a bandwidth of 1,935 Gbit/s
  • Comprehensive basic capabilities
    • User-defined network with flexible subnet division and network access policy configuration
    • Mass storage, elastic expansion, and backup and restoration
    • Elastic scaling
  • Flexibility

    Similar to other types of ECSs, P3 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 P3 ECSs.

Supported Software

P3 ECSs are used in computing acceleration scenarios, such as deep learning training, inference, scientific computing, molecular modeling, and seismic analysis. If the software is required to support GPU CUDA, use P3 ECSs. P3 ECSs support the following commonly used software:

  • Common deep learning frameworks, such as TensorFlow, Spark, PyTorch, MXNet, and Caffe
  • CUDA GPU rendering supported by RedShift for Autodesk 3ds Max and V-Ray for 3ds Max
  • Agisoft PhotoScan
  • MapD
  • More than 2,000 GPU-accelerated applications such as Amber, NAMD, and VASP

Notes

  • After a P3 ECS is stopped, basic resources (including vCPUs, memory, image, and GPUs) are not billed, but its system disk is billed based on the disk capacity. If other products, such as EVS disks, EIP, and bandwidth are associated with the ECS, these products are billed separately.

    Resources will be released after a P3 ECS is stopped. If resources are insufficient at the next start, the start may fail. If you want to use such an ECS for a long period of time, do not stop the ECS.

  • If a P3 ECS is created using a private image, make sure that the Tesla driver was installed during the private image creation. If not, install the driver for computing acceleration after the ECS is created. For details, see Manually Installing a Tesla Driver on a GPU-accelerated ECS.
  • GPU-accelerated ECSs differ greatly in general-purpose and heterogeneous compute. Their specifications can only be changed to other specifications of the same instance type.
  • GPU-accelerated ECSs do not support live migration.

Computing-accelerated P2s

Overview

P2s ECSs use NVIDIA Tesla V100 GPUs to provide flexibility, high-performance computing, and cost-effectiveness. P2s 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 4 P2s ECS specifications

Flavor

vCPUs

Memory

(GiB)

Max./Assured Network Bandwidth (Gbit/s)

Max. Network PPS (10,000)

Max. NIC Queues

Max. NICs

GPUs

GPU Connection

GPU Memory (GiB)

Virtualization

p2s.2xlarge.8

8

64

10/4

50

4

4

1 × V100

PCIe Gen3

1 × 32 GiB

KVM

p2s.4xlarge.8

16

128

15/8

100

8

8

2 × V100

PCIe Gen3

2 × 32 GiB

KVM

p2s.8xlarge.8

32

256

25/15

200

16

8

4 × V100

PCIe Gen3

4 × 32 GiB

KVM

p2s.16xlarge.8

64

512

30/30

400

32

8

8 × V100

PCIe Gen3

8 × 32 GiB

KVM
P2s ECS Features
  • CPU: 2nd Generation Intel® Xeon® Scalable 6278 processors (2.6 GHz of basic frequency and 3.5 GHz of turbo frequency), or Intel® Xeon® Scalable 6151 processors (3.0 GHz of basic frequency and 3.4 GHz of turbo frequency)
  • 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
  • 14 TFLOPS of single-precision computing and 7 TFLOPS of double-precision computing
  • NVIDIA Tensor cores with 112 TFLOPS of single- and double-precision computing for deep learning
  • Up to 30 Gbit/s of network bandwidth on a single ECS
  • 32 GiB of HBM2 GPU memory with a bandwidth of 900 Gbit/s
  • Comprehensive basic capabilities
    • User-defined network with flexible subnet division and network access policy configuration
    • Mass storage, elastic expansion, and backup and restoration
    • Elastic scaling
  • Flexibility

    Similar to other types of ECSs, P2s 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 P2s ECSs.

Supported Software

P2s ECSs are used in computing acceleration scenarios, such as deep learning training, inference, scientific computing, molecular modeling, and seismic analysis. If the software is required to support GPU CUDA, use P2s ECSs. P2s 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 3ds Max and V-Ray for 3ds Max
  • Agisoft PhotoScan
  • MapD
Notes
  • After a P2s ECS is stopped, basic resources (including vCPUs, memory, image, and GPUs) are not billed, but its system disk is billed based on the disk capacity. If other products, such as EVS disks, EIP, and bandwidth are associated with the ECS, these products are billed separately.

    Resources will be released after a P2s ECS is stopped. If resources are insufficient at the next start, the start may fail. If you want to use such an ECS for a long period of time, do not stop the ECS.

  • By default, P2s ECSs created using a public image have the Tesla driver installed.
  • If a P2s ECS is created using a private image, make sure that the Tesla driver was installed during the private image creation. If not, install the driver for computing acceleration after the ECS is created. For details, see Manually Installing a Tesla Driver on a GPU-accelerated ECS.
  • GPU-accelerated ECSs differ greatly in general-purpose and heterogeneous compute. Their specifications can only be changed to other specifications of the same instance type.
  • GPU-accelerated ECSs do not support live migration.

Computing-accelerated P3snl

Overview

P3snl ECSs use NVIDIA A100 GPUs and provide flexibility and ultra-high-performance computing. P3snl ECSs have strengths in AI-based deep learning, scientific computing, Computational Fluid Dynamics (CFD), computing finance, seismic analysis, molecular modeling, and genomics. Theoretically, the FP32 is 19.5 TFLOPS, and the TF32 tensor core is 156 TFLOPS | 312 TFLOPS (sparsity enabled).

Specifications

Table 5 P3snl ECS specifications

Flavor

vCPUs

Memory

(GiB)

Max./Assured Network Bandwidth (Gbit/s)

Max. Network PPS (10,000)

Max. NIC Queues

Max. NICs

GPUs

GPU Connection

GPU Memory

(GiB)

Virtualization

p3snl.2xlarge.8

8

64

10/4

100

4

4

1 × NVIDIA A100 40GB

PCIe Gen3

1 × 40 GiB

KVM

p3snl.4xlarge.8

16

128

15/8

200

8

8

2 × NVIDIA A100 40GB

PCIe Gen3

2 × 40 GiB

KVM

p3snl.8xlarge.8

32

256

25/15

350

16

8

4 × NVIDIA A100 40GB

PCIe Gen3

4 × 40 GiB

KVM

p3snl.16xlarge.8

64

512

30/30

700

32

8

8 × NVIDIA A100 40GB

PCIe Gen3

8 × 40 GiB

KVM
P3snl ECS Features
  • CPU: 2nd Generation Intel® Xeon® Scalable 6248R processors and 3.0 GHz of basic frequency
  • Up to eight NVIDIA A100 GPUs on an ECS
  • NVIDIA CUDA parallel computing and common deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet
  • 19.5 TFLOPS of single-precision computing and 9.7 TFLOPS of double-precision computing on a single GPU
  • NVIDIA Tensor cores with 156 TFLOPS of single- and double-precision computing for deep learning
  • Up to 40 Gbit/s of network bandwidth on a single ECS
  • 40 GiB of HBM2 GPU memory with a bandwidth of 1,935 Gbit/s
  • Comprehensive basic capabilities
    • User-defined network with flexible subnet division and network access policy configuration
    • Mass storage, elastic expansion, and backup and restoration
    • Elastic scaling
  • Flexibility

    Similar to other types of ECSs, P2s 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 P3snl ECSs.

Supported Software

P3snl ECSs are used in computing acceleration scenarios, such as deep learning training, inference, scientific computing, molecular modeling, and seismic analysis. If the software is required to support GPU CUDA, use P3snl ECSs. P3snl 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 3ds Max and V-Ray for 3ds Max
  • Agisoft PhotoScan
  • MapD
  • More than 2,000 GPU-accelerated applications such as Amber, NAMD, and VASP
Notes
  • After a P3snl ECS is stopped, basic resources (including vCPUs, memory, image, and GPUs) are not billed, but its system disk is billed based on the disk capacity. If other products, such as EVS disks, EIP, and bandwidth are associated with the ECS, these products are billed separately.

    Resources will be released after a P3snl ECS is stopped. If resources are insufficient at the next start, the start may fail. If you want to use such an ECS for a long period of time, do not stop the ECS.

  • By default, P3snl ECSs created using a public image have the Tesla driver installed.
  • If a P3snl ECS is created using a private image, make sure that the Tesla driver was installed during the private image creation. If the Tesla driver has not been installed, install the driver for computing acceleration after the ECS is created. For details, see Manually Installing a Tesla Driver on a GPU-accelerated ECS.
  • GPU-accelerated ECSs differ greatly in general-purpose and heterogeneous compute. Their specifications can only be changed to other specifications of the same instance type.
  • GPU-accelerated ECSs do not support live migration.

Inference-accelerated Pi2

Overview

Pi2 ECSs use NVIDIA Tesla T4 GPUs dedicated for real-time AI inference. These ECSs use the T4 INT8 calculator for up to 130 TOPS of INT8 computing. The Pi2 ECSs can also be used for light-load training.

Specifications

Table 6 Pi2 ECS specifications

Flavor

vCPUs

Memory

(GiB)

Max./Assured Network Bandwidth (Gbit/s)

Max. Network PPS

(10,000)

Max. NIC Queues

GPUs

GPU Memory

(GiB)

Local Disks

Virtualization

pi2.2xlarge.4

8

32

10/4

50

4

1 × T4

1 × 16

-

KVM

pi2.4xlarge.4

16

64

15/8

100

8

2 × T4

2 × 16

-

KVM

pi2.8xlarge.4

32

128

25/15

200

16

4 × T4

4 × 16

-

KVM

pi2.16xlarge.4

64

256

30/30

400

32

8 × T4

8 × 16

-

KVM

Pi2 ECS Features

  • CPU: 2nd Generation Intel® Xeon® Scalable 6278 processors (2.6 GHz of basic frequency and 3.5 GHz of turbo frequency), or Intel® Xeon® Scalable 6151 processors (3.0 GHz of basic frequency and 3.4 GHz of turbo frequency)
  • Up to four NVIDIA Tesla T4 GPUs on an ECS
  • GPU hardware passthrough
  • Up to 8.1 TFLOPS of single-precision computing on a single GPU
  • Up to 130 TOPS of INT8 computing on a single GPU
  • 16 GiB of GDDR6 GPU memory with a bandwidth of 320 GiB/s on a single GPU
  • Built-in one NVENC and two NVDECs

Supported Software

Pi2 ECSs are used in GPU-based inference computing scenarios, such as image recognition, speech recognition, and natural language processing. The Pi2 ECSs can also be used for light-load training.

Pi2 ECSs support the following commonly used software:

  • Deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet

Notes

  • After a Pi2 ECS is stopped, basic resources (including vCPUs, memory, image, and GPUs) are not billed, but its system disk is billed based on the disk capacity. If other products, such as EVS disks, EIP, and bandwidth are associated with the ECS, these products are billed separately.

    Resources will be released after a Pi2 ECS is stopped. If resources are insufficient at the next start, the start may fail. If you want to use such an ECS for a long period of time, do not stop the ECS.

  • Pi2 ECSs support automatic recovery when the hosts accommodating such ECSs become faulty.
  • By default, Pi2 ECSs created using a public image have the Tesla driver installed.
  • If a Pi2 ECS is created using a private image, make sure that the Tesla driver was installed during the private image creation. If not, install the driver for computing acceleration after the ECS is created. For details, see Manually Installing a Tesla Driver on a GPU-accelerated ECS.
  • GPU-accelerated ECSs differ greatly in general-purpose and heterogeneous compute. Their specifications can only be changed to other specifications of the same instance type.
  • GPU-accelerated ECSs do not support live migration.

Inference-accelerated Pi2nl

Overview

Pi2nl ECSs use NVIDIA Tesla T4 GPUs dedicated for real-time AI inference. These ECSs use the T4 INT8 calculator for up to 130 TOPS of INT8 computing. The Pi2nl ECSs can also be used for light-workload training.

Specifications
Table 7 Pi2nl ECS specifications

Flavor

vCPUs

Memory

(GiB)

Max./Assured Network Bandwidth

(Gbit/s)

Max. Network PPS

(10,000)

Max. NIC Queues

Max. NICs

GPUs

GPU Memory

(GiB)

Local Disks

Virtualization

pi2nl.2xlarge.4

8

32

10/4

50

4

4

1 × T4

1 × 16

-

KVM

pi2nl.4xlarge.4

16

64

15/8

100

8

8

2 × T4

2 × 16

-

KVM

pi2nl.8xlarge.4

32

128

25/15

200

16

8

4 × T4

4 × 16

-

KVM

pi2nl.16xlarge.4

64

256

30/30

400

32

8

8 × T4

8 × 16

-

KVM

Pi2nl ECS Features

  • CPU: 2nd Generation Intel® Xeon® Scalable 6278 processors (2.6 GHz of basic frequency and 3.5 GHz of turbo frequency), or Intel® Xeon® Scalable 6151 processors (3.0 GHz of basic frequency and 3.4 GHz of turbo frequency)
  • Up to four NVIDIA Tesla T4 GPUs on an ECS
  • GPU hardware passthrough
  • Up to 8.1 TFLOPS of single-precision computing on a single GPU
  • Up to 130 TOPS of INT8 computing on a single GPU
  • 16 GiB of GDDR6 GPU memory with a bandwidth of 320 GiB/s on a single GPU
  • Built-in one NVENC and two NVDECs

Supported Software

Pi2nl ECSs are used in GPU-based inference computing scenarios, such as image recognition, speech recognition, and natural language processing. The Pi2nl ECSs can also be used for light-load training.

Pi2 ECSs support the following commonly used software:

  • Deep learning frameworks, such as TensorFlow, Caffe, PyTorch, and MXNet

Notes

  • After a Pi2nl ECS is stopped, basic resources (including vCPUs, memory, image, and GPUs) are not billed, but its system disk is billed based on the disk capacity. If other products, such as EVS disks, EIP, and bandwidth are associated with the ECS, these products are billed separately.

    Resources will be released after a Pi2nl ECS is stopped. If resources are insufficient at the next start, the start may fail. If you want to use such an ECS for a long period of time, do not stop the ECS.

  • Pi2nl ECSs support automatic recovery when the hosts accommodating such ECSs become faulty.
  • By default, Pi2nl ECSs created using a public image have the Tesla driver installed.
  • If a Pi2nl ECS is created using a private image, make sure that the Tesla driver was installed during the private image creation. If the Tesla driver has not been installed, install the driver for computing acceleration after the ECS is created. For details, see Manually Installing a Tesla Driver on a GPU-accelerated ECS.
  • GPU-accelerated ECSs differ greatly in general-purpose and heterogeneous compute. Their specifications can only be changed to other specifications of the same instance type.
  • GPU-accelerated ECSs do not support live migration.
Parent topic: x86 ECS Specifications