VRAM

in English: video RAM (GPU memory)

advancedwidespread

VRAM is the graphics card's dedicated, high-speed memory; in AI, it determines how large a model can fit and run on your machine.

In plain terms

VRAM is the video card's (GPU) own memory. Originally, it stored the image displayed on the screen and the graphics of games. Today, it is one of the most critical bottlenecks for running AI models.

Imagine the graphics card as a very fast chef and the VRAM as the workbench. No matter how lightning-fast the chef is, if the bench is small: anything not within arm's reach requires running to the pantry, and cooking slows down significantly.

A language model consists of billions of numbers, and these must ideally be in VRAM when answering. If the model does not fit, it either fails to start, or the program moves part of the model to slower system memory, and the response trickles out painfully slowly.

Therefore, when running AI on your own machine, the first question is usually not how fast the card is, but how many gigabytes of VRAM it has.

How it works

At startup, the program loads the model weights (the learned numbers) from storage into the graphics card memory. This accounts for most of the space requirement.
Storage requirements depend on the model size and the precision of the numbers. A 7–8 billion parameter model with standard 16-bit precision requires roughly 14–16 GB, while compressed (quantized) it can fit into 4–6 GB.
While working, the context also takes up space: the model keeps intermediate calculation results related to the previous text of the conversation in memory. The longer the document you give the model, the more VRAM is needed in addition to the weights.
Response speed is largely determined by how quickly the GPU can access the weights. From its own memory, it does this at very high bandwidth. What does not fit goes into system memory, and that part is typically computed by the CPU or has to be sent to the card over a slow link, so it is much slower.
In data centers, a large model is distributed across the memory of multiple graphics cards, so the cards' memory effectively adds up.

An everyday example

You download a free language model to run offline on your laptop, but the program displays 'out of memory' or starts but responds very slowly. The 12 GB model does not fit on your 8 GB VRAM graphics card, so you must choose a smaller or more heavily compressed version.

Who is affected

developerThe VRAM size determines which model you can run locally or fine-tune, and what context size the coding agent can work with. → Code

entrepreneur, executiveIf a company runs AI on its own machine for data protection reasons, the required VRAM determines most of the hardware cost.

parent, student, teacherWhen buying a computer, it is worth knowing that for AI tasks, the video card's memory matters more than the flashy model name. → Education & Family

industrial worker, engineerThe performance of camera-based inspection on the production line or locally running models is also limited by the built-in device's memory. → Industry

A common misconception

MythIf VRAM is scarce, adding more regular RAM to the machine solves it just as well.

In realitySystem memory cannot replace that of the graphics card. The model may start, but parts forced outside the GPU will run much slower. Moreover, VRAM is almost always soldered to the card and cannot be upgraded afterward.

Limits and risks

VRAM cannot be expanded later, so you must decide how much you will need at the time of purchase; more memory requires a new card.
High VRAM does not automatically mean a fast machine: it shows what fits, not the speed at which it runs.
Compression (quantization) reduces space requirements, but with strong compression, the quality of the model's responses deteriorates.
Large-memory cards are expensive and in high demand; AI industry demand sometimes drives up prices for memory and graphics cards.

Similar, but not the same

RAM (system memory) – RAM is the processor's general working memory. VRAM belongs to the graphics card and is designed for the GPU to access it directly and quickly.
Storage (SSD, hard drive) – The long-term storage permanently preserves the downloaded model, while the VRAM only holds it during operation and clears upon shutdown.
Unified Memory – In some machines (such as Apple's own chip-based Macs or devices with integrated graphics), the processor and graphics unit share common memory. There is no separate VRAM; most of the shared memory can also be used by the model.

Related concepts

Broader

computer memory, AI infrastructure, Hardver

Narrower

HBM (high-bandwidth memory), GDDR memory

On the edge

GPU (grafikus processzor), Quantization, Helyben futtatott modell, Kontextusablak, Inference

A brief history

The name VRAM originally referred to a dual-port memory chip type designed specifically for screen handling, released in the 1980s. The chip type itself has disappeared, but the name remains, and today it generally refers to the memory of a graphics card.

Where things stand

At the end of 2026, VRAM is one of the most important purchasing criteria for locally run AI. Home cards typically offer 8–32 GB, while data center accelerators offer several times that. Meanwhile, the AI industry's memory demand is causing shortages and price rises across the whole market.

Status: 2026 November

Also known as

VRAM · video RAM · video memory · video memory · GPU memory · GPU memory · graphics memory · graphics card memory

Sources

Wikipedia: Video random-access memory

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.