Elastic GPU

Powerful parallel computing capabilities based on GPU technology.

Elastic GPU Service (EGS) is a GPU-based computing service ideal for scenarios such as deep learning, video processing, scientific computing, and visualization. EGS solutions use the following GPUs: AMD FirePro S7150, NVIDIA Tesla M40, NVIDIA Tesla P100, NVIDIA Tesla P4, and NVIDIA Tesla V100.

Benefits

Deep Learning
Online deep learning training and inference services, image recognition, content identification, and voice recognition
Video Processing
HD media coding, 4K/8K HD live, video conferencing, and source film repair
Scientific Computing
Video rendering, collision simulation, computational finance, genetic engineering, and climate prediction
Visualization
Engineering design, non-linear editing, and distance education applications

Features

  • Computing capabilities

    Extremely powerful computing capabilities of various GPUs


    GA1 instance

    A GA1 instance can provide a maximum of four AMD Fire Pro S7150 GPUs, 56 vCPUs, and 160 GB of memory. It has 32 GB of GPU memory and 8192 cores that work in parallel, and delivers up to 15 TFLOPS of single-precision, and 1 TFLOPS double-precision, floating-point performance.


    GN4 instance

    A GN4 instance can provide a maximum of two NVIDIA Tesla M40 GPUs, 56 vCPUs, and 96 GB of memory. It has 24 GB of GPU memory and 6000 cores that work in parallel, and delivers up to 14 TFLOPS of single-precision floating-point performance.


    GN5 instance

    A GN5 instance can provide a maximum of eight NVIDIA Tesla P100 GPUs, 56 vCPUs, 480 GB of memory, and 128 GB of GPU memory. It delivers up to 74.4 TFLOPS of single-precision floating-point performance. This helps achieve large-scale parallel floating-point computation performance required in deep learning and other general-purpose GPU computation scenarios. A GN5 instance also provides up to 37.6 TFLOPS of double-precision floating-point performance to deliver high computing performance required in scenarios such as scientific computing.


    GN5i instance

    A GN5i instance can provide a maximum of two NVIDIA Tesla P4 GPUs, 56 vCPUs, and 224 GB of memory. It has 16 GB of GPU memory and delivers up to 11 TFLOPS of single-precision floating-point performance and 44 TOPS INT8 of computing capability.


    GN6 instance

    A GN6 instance can provide a maximum of eight NVIDIA Tesla V100 GPUs, 88 vCPUs, and 256 GB of memory. It has 128 GB of GPU memory. Using Tensor Cores, a GN6 instance can provide up to 1000 TFLOPS of deep learning computing capability, and a single-precision floating-point performance of 125.6 TFLOPS. This helps achieve large-scale parallel floating-point computation performance required in general-purpose GPU computation scenarios. A GN6 instance also provides up to 62.4 TFLOPS of double-precision floating point performance to deliver high computing performance required in scenarios such as scientific computing.

  • Extraordinary general network performance

    The excellent network performance delivered by EGS maximizes computing and rendering performance for a wide range of complex computational scenarios.


    Satisfying network performance over computation nodes

    An Elastic GPU instance supports up to 2,000,000 PPS and 25 Gbit/s of internal network bandwidth to provide optimal network performance required by computation nodes.


    Powerful storage performance

    Elastic GPU instances have a high-speed local cache, and can be attached with ultra cloud disks or SSD cloud disks. This ensures high availability of data and maximizes the computation and rendering performance.

  • Multiple payment methods

    You can choose the payment method that best suits your needs.


    Pay yearly

    Pay for instance use on a yearly basis to maximize your discount benefit.


    Pay monthly

    Pay for instance use on a monthly basis to maintain reasonable costs during each payment while also enjoying relatively low hourly price for instance use.


    Pay hourly

    Pay for instance use on an hourly basis to satisfy your temporary need of compute resources. You are billed for the lowest cost each time.


    Request spot instances

    You can request spot instances and receive up to a 90% discount.

  • Running high-performance NVMe drives (special feature of GA1 and GN5 instances)

    High-performance NVM Express drives running on GPU instances


    High-performance NVMe drives running on GPU instances

    Highly reliable cloud storage that is based on three-copy redundancy can be attached to GA1 instances. Additionally, NVMe drives with up to 1.4 TB capacity can also run on GA1 instances. These NVMe drives can handle 230,000 IOPS with an I/O latency of about 200 μs, and provide up to 1900 Mbit/s of read bandwidth and 1100 Mbit/s of write bandwidth. (The instance performance is tested with random 240,000 reads and an IO depth of 12.)

Common Scenarios

  • Online rendering in the cloud (GA1)
  • General-purpose GPU computation (GN4)
  • Outstanding computation acceleration (GN5)
  • Deep learning inference capabilities (GN5i)
Online rendering in the cloud (GA1)

Online rendering in the cloud

Online rendering using Cloud Desktop

You can quickly access a GA1 instance using Cloud Desktop to experience richer visual and manipulation renderings. You can also use the Remote Desktop Protocol (RDP) to achieve real-time online rendering and graph editing. By using RDP, you can access a GA1 instance from anywhere and perform rendering and graph editing work using multiple types of devices. Data is stored using Network Attached Storage (NAS) or Alibaba Cloud Object Storage Service (OSS). You can pull data from your internal network at any time, which ensures data security. In workplaces, Express Connect and NAT Gateway can be used to improve network experiences and reduce costs.

Currently, GA1 instances only support Windows Server 2008 R2 (64-bit), Windows 7 (64-bit), CentOS 7.3 (64-bit), and Ubuntu 16.04 (64-bit). Support for Windows Server 2016 and Windows 10 is coming soon.

Benefits

  • Visualized instances

    With the powerful computing performance of GA1 instances, you can complete online editing from anywhere.

  • Service integrations

    GA1 instances can be integrated with services such as Express Connect, NAT Gateway, OSS, and NAS.

Integrations and Configurations

General-purpose GPU computation (GN4)

Excellent acceleration capability suitable for general-purpose GPU computation

Acceleration engine provided for deep learning

A GN4 instance is based on NVIDIA's Maxwell M40 GPU and provides up to 14 TFLOPS of single-precision floating-point performance. This helps achieve large-scale parallel floating-point computation performance required in deep learning and other general-purpose GPU computation scenarios. GN4 instances can be seamlessly integrated into an elastic computing ecosystem to provide solutions that are ideal for either online or offline computation scenarios. Additionally, integrating Container Service into your workflow can help simplify deployment and O&M, and provide resource scheduling services.

Benefits

  • Elastic expansion

    GN4 instances can interwork with Auto Scaling and Server Load Balancer to achieve elastic expansion.

  • Fast deployment

    Using Container Service can speed up service deployment, O&M, and resource scheduling.

Integrations and Configurations

Outstanding computation acceleration (GN5)

Outstanding floating-point computation acceleration capability

Outstanding computation acceleration performance

A GN5 instance is based on NVIDIA Tesla P100 GPU and provides up to 74.4 TFLOPS of single-precision floating-point performance. This helps achieve large-scale parallel floating-point computation performance required in deep learning and other general-purpose GPU computation scenarios. A GN5 instance also provides up to 37.6 TFLOPS of double-precision floating-point performance to deliver high computing performance required in scenarios such as scientific computing. GN5 instances support the GPUDirect P2P technology. In this way, GPUs can directly communicate with each other by using PCI buses, greatly reducing inter-GPU communication latency. GN5 instances can be seamlessly integrated into an elastic computing ecosystem to provide solutions that are ideal for either online or offline computation scenarios.

Additionally, making full use of Container Service can help simplify deployment and O&M, and provide resource scheduling services. The Image Market provides a GN5 instance image that is equipped with an NVIDIA GPU driver and a deep learning framework, which simplifies deployment.

Benefits

  • Elastic expansion

    GN5 instances can interwork with Auto Scaling and Server Load Balancer to achieve elastic expansion.

  • Fast deployment

    Using Container Service can speed up service deployment, O&M, and resource scheduling.

Integrations and Configurations

Deep learning inference capabilities (GN5i)

Extraordinary deep learning inference capabilities

Optimal deep learning inference capabilities

A GN5i instance is based on NVIDIA Tesla P4 GPU and provides up to 11 TFLOPS of single-precision floating-point performance and 44 TOPS INT8 of computing capability that are ideal for deep learning scenarios, especially for inference. Additionally, a single GPU only consumes 75 W of power while maintaining a high-performance output. GN5i instances can be seamlessly integrated into an elastic computing ecosystem to provide solutions that are ideal for either online or offline computation scenarios. Additionally, making full use of Container Service can help simplify deployment and O&M, and provide resource scheduling services. The Image Market provides a GN5i instance image that is equipped with an NVIDIA GPU driver and a deep learning framework, which simplifies deployment.

Benefits

  • Elastic expansion

    GN5 instances can interwork with Auto Scaling and Server Load Balancer to achieve elastic expansion.

  • Fast deployment

    Using Container Service can speed up service deployment, O&M, and resource scheduling.

Integrations and Configurations