How to Increase VRAM in 2026: What Actually Works on PC, Laptop and Mac

Search “how to increase VRAM” and most results give you a BIOS menu and a registry edit, as if graphics memory were a setting. Mostly it isn’t. Whether you can get more VRAM depends on where the memory physically lives. On a graphics card it’s soldered to the board. On an integrated GPU or an Apple Silicon Mac it’s system RAM, and there you often can raise the share the GPU is allowed to use. This guide covers what actually works on each kind of hardware in 2026, what’s a myth, and how to make AI models and games need less memory in the first place.

Quick answer

What VRAM does

VRAM (video RAM) is the memory your GPU works from. Games keep textures, geometry and frame buffers in it. AI models keep their weights and working state in it: the KV cache for a language model, activations for an image model. VRAM decides what fits. When a workload needs more than you have, it either refuses to load or spills into much slower system memory.

VRAM doesn’t decide how fast things run once they fit. That comes down to the GPU’s compute and memory bandwidth. VRAM vs RAM explains the difference.

Dedicated vs shared GPU memory in Task Manager

Windows Task Manager shows two figures under Performance → GPU. Microsoft’s DirectX team defines them this way. Dedicated GPU memory is memory reserved exclusively for the GPU; on a graphics card, that’s its VRAM. Shared GPU memory is ordinary system RAM that either the GPU or the CPU can use as needed. Windows caps it: the GPU may use at most half of physical memory at any moment, so on a 16GB PC the GPU can borrow up to 8GB.

That’s why a PC with a 4GB card can report 12GB of “total graphics memory”. The extra 8GB is system RAM the card can fall back on, and it’s far slower than the card’s own VRAM.

Graphics cards: you can’t add VRAM

On a desktop or laptop graphics card, the memory chips are soldered to the board next to the GPU. No setting, driver or registry value adds any. To get more VRAM you need a card with more of it. Our AI GPU comparison ranks current cards by VRAM, and the best GPU for AI guide covers what each tier is good for.

What you can change on NVIDIA is what happens when VRAM runs out. Recent Windows drivers can let CUDA applications spill into system memory instead of failing with an out-of-memory error. The control is CUDA – Sysmem Fallback Policy in the NVIDIA Control Panel. NVIDIA’s support article covers setting it to Prefer No Sysmem Fallback for Stable Diffusion, and you restart the application afterwards. Fallback lets an oversized job finish rather than crash, but it runs much slower. Turning it off means you see the out-of-memory error straight away instead of a silent slowdown.

Intel integrated graphics: a BIOS limit, and a registry myth

Intel’s support article is direct about this: the integrated GPU “does not have a dedicated memory; it uses shared memory that will be allocated automatically.” On some systems the BIOS has a setting, usually listed as VRAM or DVMT (Dynamic Video Memory Technology), with options such as 128MB, 256MB or maximum. Intel notes that this only limits the maximum the GPU can take. You can’t preset VRAM to a specific value, and many laptops don’t expose the setting at all.

The registry tweak in most “increase VRAM” guides creates a value called DedicatedSegmentSize under the Intel graphics key. It changes how much “dedicated” memory Windows reports. It doesn’t move any memory, because on integrated graphics all graphics memory is system RAM to begin with. It can help an older game that refuses to start when it sees too little dedicated memory. It doesn’t make anything faster, and it doesn’t let a larger AI model load.

The change that genuinely helps an Intel integrated GPU is more system RAM, running in dual channel. The GPU can then borrow more, and borrow it faster.

AMD Ryzen: Variable Graphics Memory

AMD calls this Variable Graphics Memory (VGM): “a BIOS-level feature introduced in AMD Ryzen AI 300 series processors (and beyond) which allows the user to reallocate a % of the system RAM to integrated graphics.” Unlike the Intel registry value, it really does turn system RAM into dedicated graphics memory, and AMD says the operating system treats it as virtually identical to VRAM.

1. Open AMD Software: Adrenalin Edition.
2. Go to Performance → Tuning → Variable Graphics Memory and choose an amount from the drop-down.
3. Restart. The change takes effect at boot.

The trade-off is simple: whatever you give the GPU, the CPU loses. AMD’s flagship example is the Ryzen AI Max+ 395 with 128GB, which can dedicate up to 96GB to graphics, with 112GB addressable by the GPU in total. That’s enough to load models far larger than any consumer graphics card holds. Our Strix Halo vs Apple M5 Max comparison covers that class of machine.

Older Ryzen APUs don’t have VGM. Some motherboards offer a BIOS option for integrated-graphics memory, often labelled UMA Frame Buffer Size, but the name and the available values depend on the board maker.

Apple Silicon Macs: raise the GPU wired limit

A Mac has no separate VRAM. The CPU and GPU share one pool of unified memory, and macOS caps how much of it the GPU can wire, leaving the rest for the system. You can raise that cap with a sysctl. Apple’s MLX documentation gives the command:

sudo sysctl iogpu.wired_limit_mb=N

Replace N with a figure in megabytes. It needs to be larger than the model you want to load and smaller than your total memory. Leave generous headroom for macOS, or the whole system starts swapping. The setting resets when you restart; set it to 0 to go back to the default straight away. MLX’s documentation lists macOS 15.0 or later for this. To work out a sensible value of N for your Mac and the model you want to run, use our Mac VRAM optimizer, which generates the command for you. Our best Mac for local LLMs guide covers how much memory to buy in the first place.

Laptops

A laptop follows the same rules as its GPU. If it has a dedicated graphics chip, that VRAM is fixed. If it relies on integrated graphics, you get the Intel or AMD options above, when the BIOS or driver exposes them. Two real options remain. You can add system RAM, if the laptop has slots rather than soldered memory, which gives an integrated GPU more to borrow. Or you can connect an external GPU over Thunderbolt or USB4, which adds a whole graphics card and its VRAM, at the cost of a slower connection.

Make your workload need less VRAM

If you can’t add memory, reduce what you ask of it. For local AI, this is usually where the biggest wins are.

Use a smaller quantization

Model weights dominate the memory use, and precision sets their size. For an 8B-parameter model, the weights take 16GB at BF16, 8.5GB at Q8_0 and 4.9GB at Q4_K_M. For a 14B model, those figures are 28GB, 14.9GB and 8.6GB. Q4_K_M is the usual sweet spot for local use. The VRAM calculator works out the total, including context, for a specific model.

Shorten the context, or quantize the KV cache

The KV cache grows with context length, and what a context window costs in VRAM gives the arithmetic. In Ollama, the context defaults to 4,096 tokens and is set with OLLAMA_CONTEXT_LENGTH. Ollama can also quantize the cache, provided Flash Attention is enabled:

OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve

According to Ollama’s documentation, q8_0 uses about half the memory of the default f16 with a very small loss in precision, and q4_0 about a quarter, with a larger loss. In llama.cpp, the equivalent flags are -ctk and -ctv, and -c sets the context size.

Offload some layers to system RAM

llama.cpp’s -ngl (--n-gpu-layers) sets how many layers stay in VRAM; the rest run from system RAM. A model that doesn’t fit can still run this way, but the layers on the CPU side set the pace, so expect a big slowdown. It’s a way to try a model, not to use one all day.

For games and image generation

In games, texture quality and resolution have the biggest effect on VRAM, so lower those first. In Stable Diffusion and FLUX, FP8 weights and offloading the text encoder cut peak memory sharply. Our guide to running Stable Diffusion locally has the figures for each model, and local AI video generation covers the heavier case, where offloading is often the only way to fit a model at all.

FAQ

Does adding more RAM increase VRAM?

Not on a graphics card; its VRAM is fixed. On integrated graphics it effectively does, because the GPU borrows system RAM. Windows lets the GPU use up to half of physical memory, and AMD’s Variable Graphics Memory can turn part of it into dedicated graphics memory.

Does increasing virtual memory or the page file increase VRAM?

No. The page file is disk space standing in for system RAM. It doesn’t add GPU memory, and anything that ends up there is far too slow for graphics work.

Is 8GB of VRAM enough?

For 1080p gaming and small local AI models, often yes. For larger language models or FLUX at full resolution, no. Our 8GB VRAM page lists which local LLMs actually fit, and models by VRAM tier covers every size up to 128GB.

How much VRAM do I need for local AI?

Enough for the model weights plus the KV cache for your context, plus about 1GB of overhead. The RAM and VRAM guide for local LLMs goes through it tier by tier.

Sources

• Intel: How to Adjust Dedicated Video Memory for Intel Graphics
• AMD: FAQs: AMD Variable Graphics Memory, VRAM and AI model sizes; AI Unleashed: Ryzen AI Max+
• Microsoft DirectX Developer Blog: GPUs in the Task Manager
• NVIDIA: System Memory Fallback for Stable Diffusion
• Apple MLX: mlx-lm README (wired memory limit)
• Ollama: FAQ (context length and K/V cache quantization)
• llama.cpp: server README (GPU layers, cache type and context flags)

Updated September 15, 2026. Model weight sizes are computed from measured GGUF bits per weight; they are not benchmark measurements.