How We Built a Custom Deep-Learning Rig: Learnings and Walkthrough

Building a computer with multiple GPUs, especially four NVIDIA GPUs, is a powerful solution for tasks requiring immense computational power such as machine learning, data analysis, 3D rendering, and high-end gaming. This guide will walk you through the process, from selecting the right components to assembling and configuring your system.

1. Planning and Component Selection

GPUs

NVIDIA GPUs: Choose the specific model of NVIDIA GPUs based on your needs. Popular choices include the NVIDIA RTX 3090, RTX 3080, or professional-grade GPUs like the NVIDIA Quadro series. Ensure all four GPUs are the same model to avoid compatibility issues.

Blower Cards: For multi-GPU setups, blower cards are essential. A blower card has a single fan that draws in air and exhausts it out the back of the case. This design prevents hot air from being recirculated inside the case, ensuring each GPU receives cooler air and maintaining an optimal operating temperature.

CPU

Processor: A high-performance CPU is essential to avoid bottlenecks. Intel’s Core i9 or AMD’s Ryzen 9 series are excellent choices. Ensure your CPU supports multiple PCIe lanes to accommodate four GPUs efficiently.

Motherboard

Motherboard: Select a motherboard with multiple PCIe x16 slots. Models like the ASUS ROG Rampage or MSI MEG series offer the required slots and spacing for multiple GPUs. Ensure the motherboard supports the total power and heat output of your GPUs.

RAM

Memory: Install at least 32GB of RAM, with 64GB or more being ideal for heavy computational tasks. Look for high-speed DDR4 or DDR5 memory that is compatible with your motherboard.

Storage

Storage: Use a fast NVMe SSD for your operating system and primary applications. Additional SSDs or HDDs can be added for extra storage.

Power Supply

Power Supply Unit (PSU): With four GPUs, you’ll need a PSU that can handle the power demand. A 1600W PSU or higher is recommended. Choose a high-efficiency, modular PSU to manage cables effectively.

Cooling

Cooling System: Effective cooling is crucial. Consider liquid cooling solutions for both the CPU and GPUs, along with additional case fans. A case with good airflow is essential.

Case

Case: A large tower case, such as the Corsair Obsidian Series or Cooler Master Cosmos, will provide the necessary space and airflow for multiple GPUs and other components.

2. Assembly

Step 1: Prepare Your Workspace

Ensure you have a clean, static-free environment to work in. Gather all tools, such as screwdrivers, and anti-static wrist straps.

Step 2: Install the CPU and RAM

Place the CPU into the motherboard socket, secure it, and attach the CPU cooler. Install the RAM sticks into the appropriate slots.

Step 3: Mount the Motherboard

Place the motherboard inside the case and secure it with screws. Connect the necessary power cables from the PSU to the motherboard.

Step 4: Install the GPUs

Insert each NVIDIA GPU into a PCIe x16 slot on the motherboard. Secure them with screws and connect the necessary power cables from the PSU to each GPU.

Step 5: Connect Storage

Install your SSDs and HDDs in the designated bays and connect them to the motherboard via SATA or NVMe connections.

Step 6: Connect Cooling Systems

Install additional case fans and connect them to the motherboard. If using liquid cooling, set up the radiators and connect the pumps to the appropriate headers on the motherboard.

Step 7: Power On and Test

Connect the PSU to the power outlet and power on the system. Ensure all components are detected and functioning correctly. Enter the BIOS to check if all GPUs and other hardware are recognized.

3. Software Configuration

Step 1: Install the Operating System

Install your preferred operating system, such as Windows or Linux. For multi-GPU setups, Linux is often preferred due to better support for high-performance computing.

Step 2: Install Drivers

Download and install the latest drivers for your NVIDIA GPUs from the NVIDIA website. Ensure the drivers are properly configured for multi-GPU support.

Step 3: Configure Software

Install any necessary software for your specific tasks. For example, install CUDA and cuDNN for machine learning applications or specific rendering software for 3D tasks.

Step 4: Optimize Settings

Optimize the settings for your GPUs using tools like NVIDIA Control Panel. Enable SLI (Scalable Link Interface) if supported and needed for your applications.

4. Maintenance and Troubleshooting

Regular Maintenance

    •    Keep your system clean: Regularly dust out your system to maintain airflow and cooling efficiency.
    •    Update software: Keep your drivers and operating system up to date to ensure optimal performance and security.
    •    Monitor performance: Use monitoring tools to keep an eye on the temperature and performance of your GPUs and other components.

Troubleshooting

    •    Check connections: Ensure all cables and components are securely connected.
    •    BIOS settings: Occasionally, check your BIOS settings to ensure all hardware is recognized and configured correctly.
    •    Driver issues: Reinstall or update drivers if you experience performance issues or crashes.

Building a computer with four NVIDIA GPUs is a significant investment but provides unparalleled performance for demanding tasks. With careful planning, the right components, and proper maintenance, you can create a powerhouse system that meets your computational needs.