Think of your computer's memory like a workspace. While the Central Processing Unit (CPU) is the worker doing calculations, it needs a place to temporarily hold the instructions and data it's actively working on. This primary workspace is called Random Access Memory (RAM), often referred to simply as system memory.

RAM is a type of electronic storage that is much faster than long-term storage devices like Solid State Drives (SSDs) or Hard Disk Drives (HDDs). When you launch a program or open a file, the necessary data is loaded from your slower storage drive into the much faster RAM so the CPU can access it quickly. This speed is essential for a smooth computing experience.

The Role of RAM in Your System

RAM serves several important functions:

Operating System: It holds parts of your operating system (like Windows, macOS, or Linux) that need to be readily available.
Applications: When you run software, like a web browser, a code editor, or the application you use to interact with an AI model, that software code is loaded into RAM.
Active Data: Data that applications are currently using, such as the text in a document you're editing or intermediate results from calculations, resides in RAM.

The key characteristic of RAM is that it's volatile. This means it only holds data while the computer has power. When you turn off your computer, everything stored in RAM disappears. That's why you need to save your work to a persistent storage device like an SSD or HDD.

Basic relationship between CPU, RAM, GPU, VRAM, and Storage. Data moves from persistent storage to RAM for general processing, and often into VRAM for GPU-intensive tasks like running LLMs.

RAM Capacity and AI Models

RAM is measured in Gigabytes (GB). Common amounts in modern computers range from 8GB to 16GB for basic tasks, 32GB for more demanding use, and 64GB or more for high-performance workstations.

How much RAM do you need for working with Large Language Models? While the LLM parameters themselves are often loaded into the specialized VRAM of a GPU (which we'll cover next), the system RAM still plays a supporting role.

System Overhead: You need enough RAM for your operating system and any background processes.
Application Software: The application used to run or interact with the LLM (like Python environments, frameworks, or user interfaces) needs RAM.
Data Handling: Data being prepared to send to the model, or received from it, might temporarily use system RAM.
Model Overflow (Less Ideal): On systems without a dedicated GPU or with insufficient VRAM, parts of the LLM might sometimes be loaded into system RAM. This is much slower than using VRAM and significantly impacts performance, but it highlights that system RAM can be a factor in model loading under certain constraints.

For typical LLM inference (using a model, not training it), the amount of system RAM is usually less restrictive than the amount of VRAM. However, having insufficient RAM (e.g., trying to run demanding software on a system with only 4GB or 8GB) can still cause slowdowns or prevent applications from running correctly, regardless of the AI task.

RAM Speed and Latency

Besides capacity (how much data it can hold), RAM also has speed (how fast data can be transferred) and latency (the delay before a transfer begins). These are measured in Megahertz (MHz) or Megatransfers per second (MT/s) and timings (like CL). While faster RAM can improve overall system responsiveness, for running pre-trained LLMs, the capacity of both RAM and especially VRAM is generally more impactful than minor differences in RAM speed.

In summary, system RAM is the computer's main workspace, essential for running the operating system, applications, and handling active data. While not typically the primary bottleneck for holding the LLM parameters themselves (that's usually VRAM), sufficient RAM is necessary for the overall system stability and performance required to work with AI models effectively. Next, we will look at the specialized hardware that excels at the parallel computations needed for LLMs: the GPU and its dedicated memory, VRAM.