Instance type categorization and naming-Elastic Compute Service(ECS)-阿里云帮助中心

Alibaba Cloud offers a wide range of instances to meet diverse business needs and use cases. This topic explains the relationships among instance family categories, instance families, and instance types, and describes the naming conventions for instance types.

Video introduction

Comparison of instance family categories

Alibaba Cloud classifies instances based on CPU architecture and their applicable business scenarios. The main categories include x86- and ARM-based compute-optimized servers, ECS Bare Metal Instances, high-performance computing (HPC) servers, Super Computing Cluster (SCC) servers, and heterogeneous computing servers that primarily use GPUs.

Family category	Description
x86 and ARM compute instance families	x86 compute instance families: Based on the x86 architecture, each vCPU corresponds to a hyper-thread of a processor core and provides stable performance. Common processor brands include Intel, AMD, and Hygon. Enterprise-level x86 compute instance families: Suitable for enterprise applications of various types and sizes, database systems, video encoding, and data analytics. Entry-level x86 compute instance families: Primarily designed for small to medium-sized websites or individual developers. Unlike enterprise-level families, these instances use a shared resource model. This means they cannot guarantee consistent computing performance but are available at a lower cost. ARM compute instance families: Based on the ARM architecture, each vCPU corresponds to a physical processor core and provides stable performance with dedicated resources. These instances are suitable for containers, microservices, web and application servers, high-performance computing, and CPU-based machine learning. Common processor brands include Yitian 710 and Ampere^® Altra^®.
ECS Bare Metal Instance (ebm) family	ECS Bare Metal Instances combine the advantages of physical servers and cloud servers to deliver exceptional and stable computing performance. Using Alibaba Cloud's proprietary virtualization technology, your applications can directly access the processor and memory of an ECS Bare Metal Instance without any virtualization overhead. These instances provide all processor features of a physical server, such as Intel VT-x, and physical-level resource isolation, which makes them ideal for migrating traditional non-virtualized workloads to the cloud.
High-performance computing (HPC) instance family	HPC-optimized instances are cost-effective solutions designed to improve the performance of HPC workloads and reduce operational costs at scale.
Super Computing Cluster (SCC) instance family	Super Computing Cluster (SCC) is built on ECS Bare Metal Instances and adds support for high-speed Remote Direct Memory Access (RDMA) interconnects. This significantly improves network performance and increases the acceleration ratio for large-scale clusters. SCC provides high bandwidth and low latency while retaining all the benefits of ECS Bare Metal Instances.
Heterogeneous computing instance family	Elastic GPU Service: Elastic GPU Service provides GPU-accelerated computing capabilities with on-demand provisioning and elastic scaling. As a part of the Alibaba Cloud Elastic Compute Service portfolio, it combines GPU and CPU computing power to meet the requirements of various scenarios, such as artificial intelligence, high-performance computing, and professional graphics and image processing. Elastic GPU Service significantly improves the computational efficiency of parallel computing tasks. Heterogeneous service instances: The video-trans instance type is suitable for video transcoding, image and video content processing, and frame extraction. Visual computing instances: The ebmgi6s instance family, which is based on the Alibaba Cloud SHENLONG architecture and Intel^® Server GPU, provides a secure and rapidly scalable architecture for the latest high-density cloud gaming and rendering instances.

Relationship between instance families and instance types

An instance family is a collection of instance types that share the same processor and similar use cases. Within an instance family, there are multiple instance types that differ in configurations, such as CPU and memory. An ECS instance type defines basic attributes, such as the CPU model and clock speed. However, an ECS instance is a complete computing service that consists of an instance type, block storage, an image, and a network type.

The relationship between instance families and instance types is as follows:

Note

The following diagram shows only a subset of instance families and instance types. For more information, see Instance families.

Instance type naming conventions

The instance family name follows the format ecs.<family>. The instance type name follows the format ecs.<family>.<size>. The naming convention is as follows:

ecs: The product code for Elastic Compute Service (ECS).
<family>: Consists of a family base and a family suffix.
<size>: Composed of small, large, or <nx>large, indicating the number of vCPUs. small means 1 vCPU, large means 2 vCPUs, and xlarge means 4 vCPUs. The larger the value of n in <n>, the more vCPUs the instance has. For example, 2xlarge represents 2 × 4 = 8 vCPUs, and 3xlarge represents 3 × 4 = 12 vCPUs, and so on.

x86 and ARM compute instance families

For example, ecs.g8ae.4xlarge is an instance type in the general-purpose instance family that uses an enhanced AMD CPU and has 16 vCPUs (4 × 4). Based on the standard 1:4 vCPU-to-memory ratio for general-purpose families, this instance type has 64 GiB of memory.

Naming differences primarily stem from variations in instance families. The table below details the components of instance family names.

Family base (lowercase letters + digits)

Family suffix

Lowercase letter

Numbers

Lowercase letter

An abbreviation indicating the performance domain of the instance family.

c: Compute-optimized

CPU-to-memory ratio of 1:2. Suitable for databases, web servers, high-performance scientific and engineering applications, game servers, data analytics, batch computing, video encoding, and machine learning.
g: General-purpose

CPU-to-memory ratio of 1:4. Suitable for general internet applications, databases, websites, Java application services, game services, and search advertising.
u: Universal compute

U instances run on various server platforms with CPU-to-memory ratios of 1:1, 1:2, 1:4, or 1:8. Ideal for price-sensitive enterprise customers running small, medium, or large enterprise applications, websites, application servers, small-to-medium database systems, caching, and search clusters.
r: Indicates memory-optimized (RAM).

CPU-to-memory ratio of 1:8 (some exceptions apply). Suitable for in-memory databases, data analytics and mining, distributed in-memory caches (Redis), big data applications (Kafka, Elasticsearch, etc.), and general enterprise applications requiring large memory capacity (e.g., Java).
re: Memory-enhanced
hf (c/g/r): High frequency

CPU-to-memory ratios of 1:2, 1:4, or 1:8. Suitable for massively multiplayer online games, HPC, and medium-to-large database systems.
i: Local SSD

CPU-to-memory ratios of 1:4 or 1:8. Suitable for OLTP, high-performance relational databases, NoSQL databases (e.g., Cassandra, MongoDB), Elasticsearch search scenarios, and EMR big data compute-storage separation.
d: Big data

CPU-to-memory ratio of 1:4 (some exceptions apply). Suitable for Hadoop MapReduce, HDFS, Hive, HBase, Elasticsearch, Kafka, and other big data computing, storage, search, and log processing scenarios.
s: Stands for shared.
- t: Indicates a burstable type.
- e: Economy

Typically indicates the release sequence among similar instance families. A higher digit represents a newer generation with better cost performance.

Examples: 8, 7, 6, 5, etc.

Generally describes additional characteristics of the instance family.

y: Uses Alibaba Cloud’s self-developed Yitian 710 ARM CPU
a: Uses AMD CPU
ae: AMD enhanced
i: Uses Intel CPU
h: Uses Hygon processor
re: RDMA enhanced
se: Storage enhanced
ne/nex: Network enhanced
t: Security enhanced (TPM)
p: Persistent memory
g: General-purpose
r: Indicates memory-optimized (RAM)
c: Compute-intensive

Note

Special instance types (e.g., ecs.e-c1m4.xlarge) use suffixes like c1m4 to indicate a vCPU-to-memory ratio of 1:4.

c: core (vCPU)
m: memory

Heterogeneous computing, ECS Bare Metal Instances, high-performance computing, and Super Computing Cluster (SCC) instance families

For example, ecs.ebmgn7ix.32xlarge is an instance type in the GPU-accelerated compute-optimized ECS Bare Metal Instance family. It provides an enhanced AMD CPU with 128 vCPUs (32 × 4) and NVIDIA A10 GPUs with 24 GB of GPU memory. The "7" in the instance type name indicates that the GPU is based on the Ampere architecture.

Important

HPC instance families provide only physical cores and do not support hyper-threading. Therefore, they are measured in physical cores instead of vCPUs. For example, ecs.hpc8i.32xlarge has 64 physical cores.

Naming differences primarily stem from variations in instance families. The table below details the components of instance family names.

Family base (lowercase letters)

Family suffix (lowercase letters + digits)

gn: Compute-optimized instances with NVIDIA GPUs
vgn: Dedicated instances accelerated by NVIDIA GRID vGPU
sgn: Shared instances accelerated by NVIDIA GRID vGPU
gi: Compute-optimized instances with Intel GPUs
f: FPGA compute-optimized instances
ebm (c/g/r/gn/hf): ECS Bare Metal Instances
scc (c/g/h/gn/hf): Super Computing Cluster
hpc: High-performance computing instances

6v: "6" indicates the Volta or Turing microarchitecture. "v" denotes a V100 GPU with 16 GB of GPU memory.

Example: gn6v indicates a compute-optimized instance that uses the Volta or Turing microarchitecture, has 16 GB of GPU memory, and is equipped with an NVIDIA V100 GPU.
6e: "6" indicates the Volta or Turing microarchitecture. "e" (extended) denotes a second-generation V100 GPU with 32 GB of GPU memory.

Example: gn6e indicates a compute-optimized instance that uses the Volta or Turing microarchitecture, has 32 GB of GPU memory, and is equipped with an NVIDIA V100 GPU.
6i: "6" indicates the Volta or Turing microarchitecture. "i" (inference) denotes a T4 GPU.

Example: gn6i indicates a compute-optimized instance that uses the Volta or Turing microarchitecture and is equipped with an NVIDIA T4 GPU.
6s: "6" indicates the Volta or Turing microarchitecture. "s" denotes a sixth-generation SG-1 chip.

Example: ebmgi6s indicates a visual computing instance with an Intel^® Server GPU and a sixth-generation SG-1 chip.
7: Indicates the Ampere microarchitecture.
7i: "7" indicates the Ampere microarchitecture. "i" (inference) denotes an A10 GPU with 24 GB of GPU memory.
7e: "7" indicates the Ampere microarchitecture. "e" denotes a version with more GPU memory. For example, a V100 GPU with 32 GB of memory compared to one with 16 GB.
7s: "7" indicates the Ampere microarchitecture. "s" denotes a seventh-generation A30 GPU.

Note

Some special instance types, such as ecs.gn7i-c8g1.2xlarge, use suffixes such as c8g1 to indicate a vCPU-to-GPU ratio of 8:1.

c: core (vCPU)
g: GPU

Instance type metrics

Metric	Description
Processor	The physical CPU model of the ECS instance. Performance varies across processors: Base CPU frequency: Also known as core frequency, this is the standard operating frequency of the processor without overclocking or special optimization. It is the typical clock speed specified by the processor designer for reliable operation under typical workloads. Turbo CPU frequency: The maximum clock frequency the processor can temporarily achieve when needed.
vCPU	For x86-based instance types, each vCPU corresponds to a hyper-thread of a processor core. For ARM-based instance types, each vCPU corresponds to a physical processor core, offering stable performance with dedicated resources.
Burstable performance	Average baseline CPU performance: The consistent CPU performance an instance can sustain. CPU credits/hour: After startup, an instance consumes CPU credits to maintain performance while earning credits at a fixed rate determined by its instance type. See the CPU credits/hour column in the instance type metrics table for the total CPU credits earned per hour across all vCPUs. Maximum CPU credit balance: The total CPU credits a burstable instance can earn in 24 hours. Credits expire after 24 hours, maintaining a dynamic balance. Since the credit earning rate is fixed for a given instance type, the credit balance has an upper limit.
Memory	Memory: Stores and retrieves data that the server can access quickly. Memory acts as temporary storage during program execution and data processing. Data in memory is volatile—it is lost when the server shuts down or restarts. Persistent memory: Can be used as either memory or local storage. In instance type metric tables that support persistent memory, it appears alongside regular memory. Supported usage modes depend on the instance type. For more information, see Configure and use persistent memory. Encrypted memory: Provided via Intel^® SGX technology. In supported instance type metric tables, total memory includes encrypted memory. For more information, see Build an SGX confidential computing environment.
Network bandwidth	Baseline network bandwidth: The initial bandwidth allocation for a network connection, representing the maximum data throughput under normal conditions. Choose a baseline bandwidth configuration based on your needs. For more information, see Network bandwidth. Burstable network bandwidth: The maximum short-term data transfer rate an instance can achieve beyond its baseline bandwidth, enabled by network burst credits. Supported starting with sixth-generation and later instance families. Burstable bandwidth leverages idle resources and is not covered by an SLA. If your workload requires guaranteed bandwidth, select an instance based on its baseline network bandwidth. For more information, see Network bandwidth. Full-duplex network bandwidth: Supported starting with seventh-generation and later instance families. Under full-duplex bandwidth, both inbound and outbound directions can simultaneously achieve the instance’s specified bandwidth (see bandwidth parameters in the instance description) without interfering with each other. You can transmit and receive data at full rates concurrently. For more information, see Internal bandwidth. Note All instance metrics are validated in pure forwarding test environments. Actual performance may vary due to workload type, packet size, connection duration, image version, networking model, and other factors. Conduct stress tests to evaluate performance before selecting an instance type.
Packet forwarding rate (PPS)	The maximum combined inbound and outbound packet forwarding capability. For testing methods, see Network performance testing methods. Note All instance metrics are validated in pure forwarding test environments. Actual performance may vary due to workload type, packet size, connection duration, image version, networking model, and other factors. Conduct stress tests to evaluate performance before selecting an instance type.
Connections	A connection (or network session) is established between a client and server to transfer data. A unique connection is identified by a 5-tuple (source IP address, destination IP address, source port, destination port, protocol). ECS instance connection counts include connections established via TCP, UDP, and ICMP protocols. If your workload is sensitive to network concurrency, choose an instance type that explicitly specifies connection limits.
Number of queues	The maximum number of queues per NIC supported by the instance type. More queues typically enable more efficient distribution and processing of network data, reducing packet wait times and improving network performance while lowering packet loss and latency. Optimal queue configuration depends on actual network load, hardware performance, and system settings. For details, see NIC multi-queue.
Elastic Network Interfaces (ENIs)	The number of ENIs the instance type supports. Each ECS instance can attach one or more ENIs. Auxiliary ENIs can be detached from one ECS instance and attached to another, enabling flexible and scalable network configurations for scenarios like multiple IP addresses, multiple NICs, and high-availability networking. For details, see Elastic Network Interface.
Number of attachable Elastic RDMA-enabled NICs (ERIs)	The number of ERIs the instance type supports. ERI (Elastic RDMA Interface) is fully compatible with RDMA communication protocols and reuses the same network as standard VPC ENIs. This lets you use RDMA functionality over your existing network without changing your network topology, achieving ultra-low latency. For details, see Elastic RDMA network.
Jumbo frame support	Indicates whether the instance type supports jumbo frames. Alibaba Cloud supports jumbo frames up to 8500 bytes, allowing Ethernet frames with 8500-byte payloads. Larger payloads improve link utilization and network performance. For setup instructions, see Jumbo frames.
Private IPv4 addresses per ENI	The number of private IPv4 addresses supported per ENI for the instance type.
IPv6 addresses per ENI	The number of IPv6 addresses supported per ENI for the instance type.
I/O optimized	I/O optimization enhances network capabilities between instances and cloud disks, ensuring full utilization of disk storage performance. I/O-optimized instances achieve the full performance of enterprise SSDs (ESSDs) when attached.
Local storage	Local storage refers to local disks attached to the physical host (host machine) where the ECS instance resides. It is temporary block storage that cannot be created independently. The console uses binary units (GiB). Warning Data reliability for local disks depends on the physical host’s reliability and carries single point of failure risk. Do not store critical long-term business data on local disks. For more information, see Local disks.
Cloud disk bandwidth	Baseline bandwidth: The maximum sustained disk bandwidth guaranteed by SLA for each instance type. Burstable bandwidth: Higher bandwidth allowed for each instance type, but limited in duration and dependent on overall system bandwidth resources. Not covered by SLA.
Disk IOPS	Baseline IOPS: The maximum sustained disk IOPS guaranteed by SLA for each instance type. Burstable IOPS: Higher IOPS allowed for each instance type, but limited in duration and dependent on overall system I/O resources. Not covered by SLA.
vTPM	Trusted computing capability: Trusted instances run on physical servers equipped with Trusted Platform Module (TPM) or Trusted Cryptography Module (TCM) as the hardware Root of Trust (RoT) within the Trusted Computing Base (TCB), enabling trusted boot and ensuring zero tampering. At the virtualization layer, virtual TPM (vTPM) provides verification of core components during instance boot.