What Is VRAM? How to Check Yours and How Much You Actually Need

What VRAM is, in plain English

VRAM stands for video random-access memory. It is a pool of very fast memory that sits on the graphics card, millimeters from the GPU chip, connected by a bus far wider than anything between your CPU and your regular RAM. Anything the GPU needs to touch constantly lives there: textures and frame buffers in a game, timeline frames in a video editor, and the weights of a neural network when you run AI locally.

Two numbers describe it. Capacity, in gigabytes, is how much fits. Bandwidth, in gigabytes per second, is how fast the GPU can read it. Gamers mostly run into the first limit. People running large language models run into both: capacity decides whether a model loads, bandwidth decides how fast it writes.

VRAM vs RAM: what is the difference?

	RAM (system memory)	VRAM (video memory)
Where it lives	Sticks on the motherboard	Chips soldered on the graphics card
Serves	The CPU	The GPU
Technology	DDR4, DDR5	GDDR6, GDDR6X, GDDR7 (HBM on data-center cards)
Typical bandwidth	60–100 GB/s (dual-channel DDR5)	360 GB/s (RTX 3060) to 1,792 GB/s (RTX 5090)
Typical capacity	16–64 GB	8–32 GB
Upgradeable	Yes	No

The two are not interchangeable. Your PC can have 64 GB of RAM and still choke on a task that needs 12 GB of VRAM, because the GPU can only work at full speed on data that sits in its own memory. When VRAM runs out, the driver starts shuttling data across the PCIe bus to system RAM. In a game that shows up as stutter. In a local LLM it shows up as generation speed falling off a cliff.

How to check how much VRAM you have

Windows 10 and 11

1. Press Ctrl + Shift + Esc to open Task Manager.
2. Open the Performance tab and click GPU in the left column.
3. Read the Dedicated GPU memory line. That is your VRAM, shown as used / total.

Ignore "Shared GPU memory" for this purpose. It is a slice of system RAM Windows is willing to lend the GPU, and it is an order of magnitude slower. A second route: press Win + R, run dxdiag, and look at Display Memory (VRAM) under the Display tab.

NVIDIA cards, any OS

nvidia-smi --query-gpu=name,memory.total,memory.used --format=csv

This is the tool to keep open while a model is running, since it shows live usage. On AMD cards under Linux, rocm-smi --showmeminfo vram does the same job.

Mac

Apple menu → About This Mac → Memory. Apple Silicon Macs have no separate VRAM; see the unified memory section below.

GPU VRAM chart: recent NVIDIA and AMD cards

Capacities below come from the BestLLMfor hardware database (90 GPUs and chips); memory type and bandwidth are manufacturer specifications. The last column uses our model catalog: the largest mainstream open-weight model whose 4-bit weights fit with room for context.

GPU	VRAM	Memory	Bandwidth	Example model that fits (Q4)
RTX 5090	32 GB	GDDR7	1,792 GB/s	Qwen 3.6 35B-A3B at Q5 (25 GB)
RTX 5080	16 GB	GDDR7	960 GB/s	Mistral Small 3.2 24B (14 GB)
RTX 5070 Ti	16 GB	GDDR7	896 GB/s	gpt-oss 20B (13 GB)
RTX 5070	12 GB	GDDR7	672 GB/s	Qwen 3 14B (9 GB)
RTX 5060 Ti	8 or 16 GB	GDDR7	448 GB/s	Qwen 3 8B (5 GB) / gpt-oss 20B (13 GB)
RTX 4090	24 GB	GDDR6X	1,008 GB/s	Qwen 3.8 27B (16 GB)
RTX 4080 / 4080 Super	16 GB	GDDR6X	717–736 GB/s	Mistral Small 3.2 24B (14 GB)
RTX 4070 / 4070 Super	12 GB	GDDR6X	504 GB/s	Qwen 3 14B (9 GB)
RTX 4060	8 GB	GDDR6	272 GB/s	Qwen 3 8B (5 GB)
RTX 3090 / 3090 Ti	24 GB	GDDR6X	936–1,008 GB/s	Qwen 3.8 27B (16 GB)
RTX 3080	10 or 12 GB	GDDR6X	760–912 GB/s	Qwen 3 14B (9 GB), short context on 10 GB
RTX 3060	12 GB	GDDR6	360 GB/s	Qwen 3 14B (9 GB)
Radeon RX 7900 XTX	24 GB	GDDR6	960 GB/s	Qwen 3.8 27B (16 GB)
Radeon RX 9070 XT	16 GB	GDDR6	640 GB/s	gpt-oss 20B (13 GB)

One detail worth knowing before you buy used: laptop GPUs carry less VRAM than the desktop card with the same name. An RTX 4090 Laptop has 16 GB, not 24; an RTX 5070 Ti Laptop has 12 GB, not 16.

How much VRAM do you need?

For gaming

Target	Comfortable VRAM	Notes
1080p, high settings	8 GB	Still fine in 2026 for most titles; the newest AAA releases at max textures can exceed it.
1440p, high settings	12 GB	The sensible floor for a new card bought today.
4K, or heavy ray tracing	16 GB+	Frame generation and path tracing add their own memory overhead.

For local AI

A language model's footprint is roughly its parameter count times its precision: at 4-bit quantization, about 0.6 GB per billion parameters, plus the KV cache that grows with conversation length, plus 1–2 GB of overhead. That arithmetic produces clear tiers:

VRAM	What it runs fully on-GPU (4-bit)
8 GB	7B–8B models comfortably; 12B with a short context
12 GB	14B-class models
16 GB	20B–24B models (gpt-oss 20B, Mistral Small 3.2 24B)
24 GB	27B–35B models (Qwen 3.8 27B, Qwen 3.6 35B-A3B)
48 GB (two 24 GB cards)	70B models (Llama 3.3 70B needs about 40 GB)

For exact numbers per model and quantization, use the VRAM calculator or the full LLM VRAM requirements table. If you already know your tier, the rankings for 8 GB, 12 GB and 24 GB cards list what to run. The reason capacity beats raw compute for this workload is covered in why VRAM matters more than TFLOPS.

Can you increase VRAM?

On a discrete graphics card, no. The memory chips are soldered to the board and sized with the GPU. The only ways to get more are a different card, or a second card: AI runtimes such as llama.cpp, Ollama and vLLM can split one model across several GPUs, so two 24 GB cards behave like 48 GB for that purpose. Games cannot do this.

The "increase VRAM" tweaks you find online apply to two narrow cases. Integrated graphics have no VRAM of their own and borrow system RAM; the BIOS often lets you enlarge that allocation (UMA frame buffer size), which raises capacity but not speed. And Windows' shared GPU memory lets a discrete card overflow into RAM, which prevents crashes and costs performance.

The real headroom is in software: pick a lighter quantization, allocate a shorter context window, or quantize the KV cache. Those three levers routinely free several gigabytes. The mechanics are in context window vs VRAM cost.

Macs: unified memory instead of VRAM

Apple Silicon has one memory pool shared by CPU and GPU, with far more bandwidth than PC system RAM: 120 GB/s on a base M4, 400 to over 600 GB/s on Max chips. By default the GPU can claim roughly three quarters of it. In our database a 24 GB Mac counts as 16 GB usable for a model, a 64 GB Mac as 48 GB, and a 128 GB Mac as 96 GB. No consumer graphics card offers 96 GB, which is why high-memory Macs have become a serious option for very large models, even though a top NVIDIA card is faster at equal capacity.

Further reading: the Wikipedia entry on video RAM for the history of the technology, NVIDIA's GeForce RTX 50 series specifications, and the llama.cpp project, whose multi-GPU and CPU-offload options are what make the AI tiers above flexible. Every capacity and model figure on this page is also available through the BestLLMfor public API (CC BY 4.0) and our open-source MCP server, if you want to query it from your own tools.