Basic concepts of GPU servers and ECS-Elastic GPU Service(EGS)-阿里云帮助中心

Key terms for Elastic GPU Service (EGS) and its underlying ECS infrastructure.

Elastic GPU Service terms

Term	Description
GPU	A GPU contains more computing units and data pipelines than a CPU, making it well-suited for large-scale parallel computing workloads.
CUDA	Compute Unified Device Architecture (CUDA) is a general-purpose parallel computing platform developed by NVIDIA for complicated GPU computing.
cuDNN	CUDA Deep Neural Network (cuDNN) is a GPU-accelerated library of primitives developed by NVIDIA for deep neural networks.
DeepGPU	DeepGPU is a free toolkit provided by Alibaba Cloud that enhances the GPU computing capabilities of Elastic GPU Service. Use DeepGPU to accelerate AI training and inference workloads at the GPU level.
Deepytorch Training	An AI training accelerator developed by Alibaba Cloud. It provides training acceleration for traditional AI and generative AI scenarios.
Deepytorch Inference	An AI inference accelerator developed by Alibaba Cloud. It significantly improves the inference performance of PyTorch by splitting the computation graph of a model, fusing execution layers, and implementing high-performance operations.
DeepNCCL	An AI communication acceleration library developed for Alibaba Cloud's X-Dragon heterogeneous products. It is used to improve communication efficiency in AI distributed training or multi-GPU inference tasks.
DeepGPU-LLM	An inference engine for large language models (LLMs) developed by Alibaba Cloud based on Elastic GPU Service. It provides high-performance inference services for LLM tasks.
FastGPU	FastGPU is an Alibaba Cloud toolkit for AI computing that simplifies the deployment of AI training and inference tasks on Alibaba Cloud IaaS resources. FastGPU provides convenient interfaces and command lines to manage deployment workflows.
cGPU	cGPU is a container sharing technology from Alibaba Cloud that isolates virtual GPUs (vGPUs) at the kernel level. Multiple isolated containers share a single physical GPU. Use cGPU when you need to run multiple workloads on one GPU instance rather than dedicating a full GPU to each container.

ECS terms

Term	Description
ECS instance	A virtual server that includes vCPUs, memory, an operating system (OS), network configurations, and disks. EGS instances are GPU-enabled ECS instances.
ECS instance type	Instance types define the computing capacity, storage capacity, and networking performance of ECS instances. Combine an instance type with an image, Elastic Block Storage (EBS) devices, and network resources to create instances for specific workloads.
image	A template containing the OS and application initialization data required to run an ECS instance. See also: public image, custom image.
public image	A licensed base image provided by Alibaba Cloud, including Windows Server and mainstream Linux distributions.
Alibaba Cloud Linux	Alibaba Cloud Linux 2 and 3 are Alibaba Cloud OS distributions optimized for ECS instances. They provide a secure, stable, high-performance environment tuned for the Alibaba Cloud infrastructure.
custom image	An image you create or import that captures a specific system environment, application state, and software configuration. Use custom images to eliminate repeated manual setup when launching multiple instances with identical configurations.
Elastic Block Storage device	A high-performance, low-latency block storage resource that you can partition, format, and mount as a file system to meet your data storage requirements.
disk	A block-level EBS device that uses a triplicate mechanism to ensure 99.9999999% data durability for ECS instances.
local disk	A storage device located on the same physical server as the ECS instance. Local disks are cost-effective and deliver high storage I/O, but their data durability is tied to the reliability of the physical server, increasing the risk of a single point of failure (SPOF). For workloads requiring high durability, use disks instead.
snapshot	A point-in-time backup of a disk, used to back up or restore data.
security group	A virtual firewall that controls inbound and outbound traffic for the ECS instances within it.
SSH key pair	A public/private key pair for authenticating instance logons. SSH key pairs apply to Linux instances only.
Instance RAM role	An instance Resource Access Management (RAM) role grants an ECS instance temporary credentials via Security Token Service (STS), allowing it to call Alibaba Cloud APIs and manage resources without embedding static keys in the instance.
virtual private cloud (VPC)	A logically isolated private network on Alibaba Cloud, isolated from other VPCs based on tunnels. Within a VPC, you control the CIDR blocks, route tables, and gateways.
elastic network interface (ENI)	An independent virtual network interface that can be attached to or detached from an ECS instance to support flexible service scaling and migration.
launch template	A saved set of instance configuration parameters that you reuse to create ECS instances without repeating manual input.
deployment set	A grouping mechanism that distributes instances across different physical servers according to a high availability strategy, ensuring business availability and disaster recovery capabilities at the underlying layer.
dedicated host	A cloud host with physical resources reserved exclusively for a single tenant. Dedicated hosts support strict security and compliance requirements and Bring Your Own License (BYOL) scenarios.
auto provisioning group	A policy-driven group that rapidly deploys instance clusters across instance types and zones, mixing preemptible and pay-as-you-go instances to provide high stability at low cost.
tag	A key-value pair attached to a resource for grouping, filtering, and managing resources by organization, purpose, or any custom dimension.
resource group	A container that organizes resources across services and regions, with independent permission management for each group.
Cloud Assistant	An automated operations and maintenance (O&M) tool from Alibaba Cloud for running commands and transferring files on ECS instances without logging in to those instances.
system event	A scheduled or unexpected O&M event that affects the running status of an ECS instance, potentially requiring a restart, stop, or release. ECS sends advance notifications with resolution steps and event cycles so you can back up data and prepare before the event occurs.