GPU云服务器

为什么选择GPU云服务器阿里云GPU云服务器是基于GPU与CPU应用的计算服务器。GPU在执行复杂的数学和几何计算方面有着独特的优势，特别是在浮点运算、并行运算等方面，GPU可以提供上百倍于CPU的计算能力。GPU的功能特性如下：拥有大量擅长处理大规模并发计算的算术逻辑单元（Arithmetic and Logic Unit，即ALU）。能够支持多线程并行的高吞吐量运算。逻辑控制单元相对简单。下表为您介绍GPU云服务器与自建GPU服务器的区别。对比项GPU云服务器GPU自建服务器灵活性能够快速开通一台或多台GPU云服务器实例。实例规格（vCPU、内存及GPU）支持灵活变更，并且支持在线升降配。带宽升降自由。服务器购买周期长。服务器规格固定，无法灵活变更。带宽一次性购买，无法自由升降。易用性Web在线管理，简单方便。内置主流的操作系统，Windows正版激活，且支持在线更换操作系统。GPU驱动可以在购买时一并安装，方便快捷。没有在线管理工具，维护困难。需用户自备操作系统，自行安装及更换。GPU驱动需要自行购买安装。容灾备份三副本数据设计，单份损坏可在短时间内快速恢复。硬件故障事故中可快速自动恢复。用户自行搭建，使用普通存储设备，价格高昂。数据损坏需用户修复。安全性能够有效阻止MAC欺骗和ARP攻击。并防护DDoS攻击，可进行流量清洗和黑洞。享有端口入侵扫描、挂马扫描、漏洞扫描等附加服务。很难阻止MAC欺骗和ARP攻击。清洗和黑洞设备需要另外购买，价格昂贵。普遍存在漏洞挂马和端口扫描等问题。成本支持包年包月及按量付费两种购买方式，可灵活选择适合您业务场景的付费方式。按需购买，无需一次性大量投入。无法按需购买，必须为业务峰值满配。一次性投入巨大，闲置浪费严重。GPU实例规格族实例是能够为您的业务提供计算服务的最小单位，不同的实例规格可以提供的计算能力也不同。根据业务场景和使用场景，ECS实例可以分为多种实例规格族，GPU实例属于ECS实例中的一类实例规格。GPU实例在提供GPU加速能力的同时，保留了与普通ECS实例一致的使用体验。在创建ECS实例时，请选择企业级异构计算规格族群、弹性裸金属服务器和超级计算集群（SCC）实例规格族群下的GPU实例规格即可。关于GPU实例规格的更多信息，请参见实例规格族。产品优势覆盖范围广阔阿里云GPU云服务器在全球17个地域实现规模部署，覆盖范围广，结合弹性供应、弹性伸缩等交付方式，能够很好地满足您业务的突发需求。超强计算能力阿里云GPU云服务器配备业界超强算力的GPU计算卡，结合高性能CPU平台，单实例可提供高达1000 TFLOPS的混合精度计算性能。网络性能出色阿里云GPU云服务器实例的VPC网络最大支持450万的PPS及32 Gbit/s的内网带宽。在此基础上，超级计算集群产品中，节点间额外提供高达50 Gbit/s的RDMA网络，满足节点间数据传输的低延时高带宽要求。购买方式灵活支持灵活的资源付费模式，包括包年包月、按量付费、抢占式实例、预留实例券、存储容量单位包。您可以按需要购买，避免资源浪费。应用场景GPU云服务器适用于视频转码、图片渲染、AI训练、AI推理、云端图形工作站等场景，神行工具包（DeepGPU）为了配合GPU云服务器的计算服务增强能力，也适用于所有AI训练场景和AI推理场景。本文主要介绍GPU云服务器和神行工具包（DeepGPU）的具体应用场景。GPU云服务器应用场景视频直播转码阿里云GPU云服务器重点支持2019年天猫双11狂欢夜直播的实时视频转码，以高画质、低带宽、高分辨率、实时的综合优势服务于天猫双11狂欢夜当天直播业务4K、2K、1080P等各个分辨率的转码。具体说明如下：GPU云服务器支持高并发实时视频流5000路以上，并逐步上升到峰值6200路每分钟，且顺利度过流量洪峰。GPU云服务器参与实时家居渲染图片生成等业务，首次提供了大量算力强劲的ebmgn6v裸金属实例，支持淘宝渲染方提升几十倍的渲染性能，第一次实现秒级实时渲染，完成总计超过5000张大型家居渲染图。AI训练GPU计算型实例规格族gn6v和gn6e具有优异的通用GPU计算加速能力，适合为深度学习提供加速引擎。具体说明如下：gn6v实例配备具有16 GB显存的NVIDIA V100 GPU计算卡，gn6e实例配备具有32 GB显存的NVIDIA V100 GPU计算卡，单节点可提供高达1000 TFlops的混合精度计算能力。实例与弹性计算生态的完美结合，为在线和离线场景提供了通用的解决方案。实例搭配容器服务使用，可以简化部署和运维的复杂度，提供资源调度服务。AI推理GPU计算型实例规格族下gn6i，gn7i等实例具有优异的AI推理能力，满足了深度学习（尤其是推理）场景下的算力需求。实例与弹性计算生态的完美结合，为在线和离线场景提供了通用的解决方案。实例搭配容器服务使用，可以简化部署和运维的复杂度，并提供资源调度服务。镜像市场提供预装NVIDIA GPU驱动和深度学习框架的镜像，简化您的部署操作。

Why choose Alibaba Cloud GPU Cloud Servers? Alibaba Cloud GPU Cloud Servers are computing servers based on GPU and CPU applications. GPUs have unique advantages in executing complex mathematical and geometric calculations, especially in floating-point operations and parallel computing. GPUs can provide computing power hundreds of times that of CPUs.

The functional features of GPUs are as follows:
1. Equipped with a large number of Arithmetic Logic Units (ALUs) that excel at processing large-scale concurrent computing;
2. Capable of supporting high-throughput multi-threaded parallel operations;
3. The logic control unit is relatively simple.

The following table introduces the differences between GPU Cloud Servers and self-built GPU servers:

| Comparison Item | GPU Cloud Server | Self-built GPU Server |
|--------------------------|----------------------------------------------------------------------------------|---------------------------------------------------------------------------------------|
| Flexibility | Can quickly provision one or more GPU cloud server instances. Instance specifications (vCPU, memory and GPU) support flexible changes, and online upgrade/downgrade is supported. Bandwidth can be adjusted freely. | Long server purchase cycle. Server specifications are fixed and cannot be changed flexibly. Bandwidth is purchased in one lump sum and cannot be adjusted freely. |
| Ease of Use | Web-based online management, simple and convenient. Pre-installed mainstream operating systems, Windows is legally activated, and online OS replacement is supported. GPU drivers can be installed at the time of purchase, which is convenient and fast. | No online management tools, difficult to maintain. Users need to prepare their own operating systems for installation and replacement. GPU drivers need to be purchased and installed by users. |
| Disaster Recovery & Backup | Three-replica data design: if one copy is damaged, it can be quickly restored in a short time. Fast automatic recovery in case of hardware failure. | Users need to build it themselves, using ordinary storage devices with high costs. Data damage requires user repair. |
| Security | Can effectively block MAC spoofing and ARP attacks. Protect against DDoS attacks, and support traffic cleaning and black hole routing. Enjoy additional services such as port intrusion scanning, malware scanning, and vulnerability scanning. | Difficult to block MAC spoofing and ARP attacks. Cleaning and black hole devices need to be purchased separately at high costs. Common issues such as malware and port scanning vulnerabilities exist. |
| Cost | Supports two purchase methods: yearly/monthly subscription and pay-as-you-go, allowing flexible selection of payment methods suitable for your business scenario. Purchase on demand without one-time large investment. | Cannot purchase on demand, must be fully configured for business peak loads. One-time huge investment, serious idle waste. |

### GPU Instance Specification Families
An instance is the smallest unit that provides computing services for your business, and different instance specifications provide different computing capabilities. According to business and usage scenarios, ECS instances can be divided into multiple instance specification families, and GPU instances are a type of instance specification under ECS instances. GPU instances provide GPU acceleration capabilities while retaining the same usage experience as regular ECS instances. When creating an ECS instance, you can select GPU instance specifications under the enterprise-level heterogeneous computing specification family, elastic bare metal server, and Super Computing Cluster (SCC) instance specification family. For more information about GPU instance specifications, see *Instance Specification Families*.

### Product Advantages
#### Wide Coverage
Alibaba Cloud GPU Cloud Servers are deployed on a large scale in 17 regions globally, with wide coverage. Combined with delivery methods such as elastic provisioning and elastic scaling, they can well meet the sudden computing demands of your business.

#### Superior Computing Power
Alibaba Cloud GPU Cloud Servers are equipped with industry-leading GPU computing cards, combined with a high-performance CPU platform. A single instance can provide up to 1000 TFLOPS of mixed-precision computing performance.

#### Excellent Network Performance
The VPC network of Alibaba Cloud GPU Cloud Server instances supports up to 4.5 million PPS and 32 Gbit/s intranet bandwidth. In addition, the Super Computing Cluster product provides an additional 50 Gbit/s RDMA network between nodes, meeting the low-latency and high-bandwidth requirements for data transmission between nodes.

#### Flexible Purchase Methods
Supports flexible resource payment models, including yearly/monthly subscription, pay-as-you-go, spot instances, reserved instance coupons, and storage capacity units. You can purchase according to your needs to avoid resource waste.

### Application Scenarios
GPU Cloud Servers are applicable to scenarios such as video transcoding, image rendering, AI training, AI inference, and cloud graphic workstations. The Shenxing Toolkit (DeepGPU) is designed to complement the computing capabilities of GPU Cloud Servers, and is also applicable to all AI training and AI inference scenarios. This article mainly introduces the specific application scenarios of GPU Cloud Servers and the Shenxing Toolkit (DeepGPU).

#### 1. Video Live Streaming Transcoding
Alibaba Cloud GPU Cloud Servers mainly supported real-time video transcoding for the 2019 Tmall Double 11 Carnival Night live broadcast, serving the live broadcast business on the day of the Tmall Double 11 Carnival Night with comprehensive advantages of high image quality, low bandwidth, high resolution, and real-time performance, including transcoding for resolutions such as 4K, 2K, and 1080P. Specific details are as follows:
GPU Cloud Servers support more than 5000 concurrent real-time video streams, gradually rising to a peak of 6200 streams per minute, and successfully passed the traffic peak. GPU Cloud Servers participated in businesses such as real-time home furnishing rendering and image generation. For the first time, a large number of powerful ebmgn6v bare metal instances were provided, supporting Taobao's rendering team to improve rendering performance by dozens of times, realizing real-time rendering in seconds for the first time, and completing a total of more than 5000 large-scale home furnishing rendering images.

#### 2. AI Training
The GPU computing instance specification families gn6v and gn6e have excellent general-purpose GPU computing acceleration capabilities, making them suitable for providing acceleration engines for deep learning. Specific details are as follows:
The gn6v instance is equipped with an NVIDIA V100 GPU computing card with 16 GB of video memory, and the gn6e instance is equipped with an NVIDIA V100 GPU computing card with 32 GB of video memory. A single node can provide up to 1000 TFLOPS of mixed-precision computing performance. The perfect combination of instances and the elastic computing ecosystem provides a universal solution for online and offline scenarios. When used with Container Service, it can simplify deployment and operation and maintenance complexity, and provide resource scheduling services.

#### 3. AI Inference
Instances such as gn6i and gn7i under the GPU computing instance specification family have excellent AI inference capabilities, meeting the computing power requirements of deep learning (especially inference) scenarios. The perfect combination of instances and the elastic computing ecosystem provides a universal solution for online and offline scenarios. When used with Container Service, it can simplify deployment and operation and maintenance complexity, and provide resource scheduling services. The Image Market provides pre-installed NVIDIA GPU drivers and deep learning framework images to simplify your deployment operations.

搜集汇总

背景与挑战

背景概述

该数据集详细介绍了阿里云GPU云服务器的核心功能，包括其超强计算能力、灵活购买方式和多场景应用（如视频转码、AI训练等），并通过对比表格突出其相较于自建服务器的优势。