Computing Ability

Highly Configured Computing Node

• Make an optimum portfolio of the CPU, memory, configuration of the local disk and the quantity of GPU to ensure an optimum computing performance of GPU.
• Each GPU Cloud Compute exclusively shares physical GPU card which ensures the stable and reliable performance of GPU.

Multiple GPU Cards

• Multiple high-performance computing cards are provided including NVIDIA Tesla V100, Tesla P40, Tesla P4, K40 with detail performance parameters.
• Each GPU Cloud Compute can be configured with 1, 2, and 4 blocks of GPU cards. The CPU, memory, and disk configurations should be promoted linearly with the number of GPUs.

Image

Integrated GPU driver

The latest GPU driver and CUDA 8.0 are integrated based on Ubuntu 16.04 image, which helps the users to solve the compatibility problems of the bottom layer and speed up the cloud service. At the same time, more images are supported one after another.

Public Image

Support more public image. You can select the released public image freely, and install the GPU driver and corresponding software according to his requirement.

Custom Image

You can generate custom image based on the system disk of the service deployed GPU Cloud Compute. Then you can create more GPU Cloud Computes through the custom image. This helps to realize rapid business replication and deployment.

Network Environment

Virtual Private Cloud (VPC)

The Virtual Private Cloud (VPC) builds a secure and isolated network environment based on Baidu AI Cloud. You can customize the network address space. The peer-to-peer high-speed connection between multiple VPCs (same city, cross city) can be realized. Through VPN/private line, a secure, customized hybrid cloud network with the user's data center can be built. The original business can be migrated to the cloud easily and safely.

100G High-performance Physical Cluster

It is designed for ultra-high-performance GPU cluster customers. It can provide 100G interconnected network between Baidu Baremetal Computes and improve the data parallelism ability between GPU clusters.

Storage System

Cloud Disk Service (CDS)

When the local disk capacity of the GPU Cloud Compute can’t meet the requirement, you can purchase CDS cloud disk to extend the storage resources.

Baidu Object storage (BOS)

When you need to upload a large amount of local data to the cloud, you can use the more cost-effective Baidu object storage BOS. When using data, you can make BOS data available as co-sharing data on multiple GPU Cloud Computes thoughBOS PS Feature.

Monitoring System

GPU Monitoring

Fine-grained status monitoring and alarm mechanism of GPU card are provided including the GPU usage rate, memory usage rate, temperature, status, etc.