UCloud GPU: Affordable GPU Compute for AI Scenarios 121

Compshare (优云智算) is a computing power platform under UCloud, offering tens of thousands of high-performance GPUs and global model API services. It primarily focuses on GPU resource rental, ai full-stack tools, and a mirror community, helping researchers, enterprises, and developers with AI model training and inference. It accelerates research output and makes usage simpler. Whether you are an individual or an enterprise, you can use it, and registering lets you claim free computing power.

Core Features of Compshare (优云智算)

• Abundant GPU resources at reasonable prices. There are tens of thousands of cards, with models including 5090 (32G), 4090 (48G/24G), 3090, A100, A800, H20, and others. Members or school/enterprise verification users get a 5% discount, with a minimum of 0.38 yuan per hour (such as 2080 8G). Each instance comes with 200G free system disk and 1.5G bandwidth, features an independent public IP, allows custom port settings, and supports adding disk space as needed. Instances are ready to use upon startup with virtually no waiting time.

• Supports both containers and virtual machines. Containers come pre-installed with common frameworks like TensorFlow and PyTorch, while virtual machines support Windows and Ubuntu operating systems, providing an experience similar to public cloud. It also accelerates academic resources, making downloads from GitHub, Hugging Face, and similar sites much faster.

• Customizable computing specifications, supporting 1 to 8 GPUs, with multi-GPU support in a single machine. Each instance features an independent public IP, facilitating external services or debugging.

• The mirror community offers many ready-to-use images, such as LLamaFactory, SD-webUI, LLM, DeepSeek, and others, allowing you to set up environments within minutes and start inference or fine-tuning quickly.

• Fast resource scheduling with instant availability upon startup and minimal queuing, ideal for performance-intensive tasks such as AI inference, deep learning training, and scientific computing.

• Round-the-clock after-sales support with technical customer service available even in the middle of the night, providing quick responses to meet the anytime needs of researchers and enterprises.

AI Full-Stack Productivity Tools

• The model API service is compatible with OpenAI and Gemini APIs. A single key can invoke popular global models, supporting LLMs, text-to-image, and text-to-video types. It responds quickly, handles concurrent requests efficiently, and integrates seamlessly with various application platforms.

• Many additional tools are available. There is a public model library that allows use without downloads, along with tools like JupyterLab and ComfyUI, supporting AI operations such as Lora training, subtitle dubbing, and speech-to-text. The platform covers everything from training to inference.

Typical Use Cases of Compshare (优云智算)

• Training and Fine-tuning: Large model pre-training, vertical-domain SFT/RLHF, multimodal alignment, and more.

• Inference and Deployment: Online services such as LLM conversations, text-to-image, digital humans, voice cloning, and recommendation systems.

• Research and Teaching: Universities or research institutions needing short-term computing power for competitions or classroom experiments.

• Industry Applications: Embodied intelligence, autonomous driving, biopharmaceuticals, financial quantitative analysis, film and TV rendering, and more.

Usage Process of Compshare (优云智算)

• Preparation: Register an account and claim benefits. You must register an account before use, and new users are eligible for computing benefits. The steps are:

• Register an Account: Visit the Compshare.cn official website (or use a special link, some of which may grant extra computing credits), enter your phone number, receive a verification code, set a password (8-24 characters containing at least three of the following: uppercase letter, lowercase letter, digit, and underscore), agree to the terms, and complete registration.

• Claim New User Benefits: Regular users can participate in promotional activities after registration. Registering with a designated link, for instance, grants 20 yuan in computing credits (equivalent to 10 hours of free 4090 cloud computing), and some activities offer up to 500 yuan in bonus credits for users who share usage blogs. School/enterprise users enjoy a 5% discount after verification, saving even more.

Core Scenario Usage Process

The Compshare (优云智算) platform serves two main purposes: AI computing (model training, deep learning, etc.) and cloud gaming (such as "Black Myth: Wukong"). The processes differ slightly.

Scenario 1: AI Computing (Using LLAMA3 Environment Setup as an Example)

• Select a GPU Instance: After logging in, navigate to the resource creation page and choose the GPU type as needed, such as 4090 24G/48G, 3090, A100, etc. The 4090 offers great value, at a minimum of 1.88 yuan per hour on hourly billing, or even cheaper with monthly plans (1.7 yuan per hour). Each instance also includes 200G free system disk and 1.5G bandwidth.

• Choose Billing Method: Options include hourly, daily, or monthly billing. Select according to your needs and avoid unnecessary long-term charges. Confirm the configuration, such as whether an independent public IP is needed, and you can set custom TCP/UDP ports for remote management convenience.

• Deploy Instances and Environment: Supports containers or virtual machines. Containers come pre-installed with frameworks like TensorFlow, PyTorch, conda, and CUDA environments, so no manual setup is required. Virtual machines support Ubuntu (with Python 3.11.7 and gcc 11.4.0 installed). For a graphical interface, you can use the platform-provided VNC visualization Docker (e.g., ImageID: compshareImage-16gb9we1sib0) for one-click setup, or manually install the XFCE4 desktop (command: apt install -y xfce4 x11vnc).

• Use Community Mirrors or Public Model Library: The platform offers over 300 AI mirrors, such as LLaMA-Factory, DeepSeek, YOLOv13, etc. With a single click, you can deploy without spending time setting up the environment. The public model library allows direct model invocation without downloads, and accelerates access to Hugging Face and GitHub, making dataset or model downloads faster.

• Connect to Instance and Operate: Multiple tools are available for remote connection. For example, use FinalShell or VSCode to connect to the terminal, and use JupyterLab or FileZilla to transfer files to the server. Access Uhost via browser, with operations similar to a local VMware virtual machine. To paste text into Linux, you can use the "press Ctrl three times consecutively → Ctrl+V to paste → press Ctrl three times consecutively again to exit" method.

• Start Computing Tasks: Once the environment is set up, you can begin model training, deep learning, and other tasks. For example, use the pre-installed LLaMA-Factory to directly start LLAMA3 training without manual CUDA configuration. To install special software (such as Java or Node.js), refer to the official documentation for self-installation.

• Resource Management and Cost Control: Be sure to manually shut down and delete resources after use. The platform disconnects instances by default upon expiration, but charges continue if the instance is not shut down (some users have accidentally left instances idle for 10 hours, incurring 26 yuan in charges). Shutting down stops billing, preventing unnecessary expenses.

• Use Academic Acceleration: The platform accelerates access to GitHub, Hugging Face, C-station, etc. No additional configuration is needed; access is automatically faster, improving efficiency.

Scenario 2: Cloud Gaming Experience (Using "Black Myth: Wukong" as an Example)

• Select a Gaming Instance: After logging in, find "Cloud Gaming" on the homepage and directly select an instance with "Black Myth: Wukong" pre-installed—no need to download or install it yourself.

• Start the Game: Click "Start," and the platform runs the game on high-performance cloud GPUs. You can play directly in your browser without worrying about local device limitations, ensuring stable frame rates with no lag or frame drops.

Key Operation Tips and Precautions

• Remote Connection and File Transfer:

• Terminal Connection: Use FinalShell by entering the instance IP, port, account, and password for remote operation. VSCode with the SSH plugin is ideal for coding and debugging.

• File Transfer: Use JupyterLab to drag and drop files directly, or use FileZilla or XFTP for FTP transfer, enabling fast transfer of large files.

• Environment Supplements: If required software (such as Java or Node.js) is not installed on the instance, you must install it yourself (refer to Oracle's official website for Java, and the official documentation for Node.js), or project execution may be affected.

• When accessing model libraries that require authorization (such as certain private models on Hugging Face), authorization must be completed on the platform in advance; otherwise, you will see an "access denied" message.

  li>

• Cost and resource management: Monitor real-time usage duration and spending, set reminders to avoid overspending. Long-term users should prioritize monthly plans for better value.

• Use "Boot Without Card": When transferring data, select the no-card mode, which doesn't use GPU compute power and only charges for storage, saving more money.