- Windows: press Ctrl+Shift+Esc → Performance → GPU and read Dedicated GPU memory. On NVIDIA cards, run
nvidia-smiin a terminal for the exact figure plus live usage. - macOS: Apple menu → About This Mac. On Apple Silicon, the “Memory” figure is unified memory shared by CPU and GPU — there is no separate VRAM.
- Linux:
nvidia-smi(NVIDIA),rocm-smi --showmeminfo vram(AMD), orglxinfo -Bon Mesa drivers. - Only dedicated VRAM determines which model fits. Plug your number into the VRAM calculator to see what you can run.
How to check VRAM, in one line per platform: on Windows, press Ctrl+Shift+Esc, open the Performance tab, select your GPU, and read Dedicated GPU memory; on macOS, open Apple menu → About This Mac; on Linux, run nvidia-smi for NVIDIA or rocm-smi for AMD. Below is each method in detail — plus what the numbers actually mean when you are deciding whether a local LLM will fit.
How to Check VRAM on Windows
Task Manager — the fastest method
- Press Ctrl+Shift+Esc to open Task Manager.
- Go to the Performance tab (on Windows 10, click More details first if you see the compact view).
- Click GPU 0 in the left sidebar. Laptops with both an integrated and a discrete GPU will also show GPU 1 — the discrete card is usually GPU 1.
- Read Dedicated GPU memory in the lower-right area. That is your physical VRAM, shown as used / total (for example, 2.1/24.0 GB).
Ignore the GPU Memory line: it adds Shared GPU memory (system RAM the GPU is allowed to borrow) to the dedicated figure, which makes an 8 GB card look like a 24 GB one. More on that distinction below.
dxdiag
Press Win+R, type dxdiag, press Enter, then open the Display tab. The Display Memory (VRAM) field shows what DirectX reports. One caveat: depending on the driver, dxdiag sometimes folds shared system memory into this number, so it can read higher than the card’s real VRAM. Treat Task Manager’s dedicated figure or nvidia-smi as the source of truth.
Settings → Display → Advanced display
Open Settings → System → Display → Advanced display, then click Display adapter properties for Display 1. The Adapter tab lists Dedicated Video Memory in MB. The exact wording differs slightly between Windows 10 and Windows 11, but the path is the same.
nvidia-smi — the exact figure
If you have an NVIDIA GPU, the driver installs a command-line tool that reports memory to the mebibyte. Open PowerShell or Command Prompt and run:
nvidia-smiThe memory column shows current usage against the true total — for example 3216MiB / 24564MiB on an RTX 4090 — and the process list at the bottom shows which programs are holding memory. For a clean, machine-readable readout:
nvidia-smi --query-gpu=memory.total,memory.used,memory.free --format=csvIf the command is not found, recent drivers place it at C:WindowsSystem32nvidia-smi.exe; older driver versions used C:Program FilesNVIDIA CorporationNVSMI.
How to Check VRAM on macOS
Click the Apple menu → About This Mac.
- Apple Silicon (M1 and later): you will see a chip name and a Memory figure such as 16 GB, 36 GB, or 128 GB. This is unified memory — a single pool shared by the CPU and GPU. There is no separate VRAM number because there is no separate VRAM.
- Intel Macs: discrete GPUs are listed with a dedicated VRAM figure, for example “Radeon Pro 5500M 8 GB”.
For more detail, open About This Mac → More Info → System Report and select Graphics/Displays, or run this in Terminal:
system_profiler SPDisplaysDataTypeOn Intel machines this prints a VRAM (Total) line; on Apple Silicon it lists the chip and GPU core count instead.
Why unified memory changes the math
On Apple Silicon the GPU can address most — not all — of system memory. macOS reserves a slice for itself, and by default the GPU’s working-set limit lands at roughly two-thirds to three-quarters of total RAM; the exact ceiling varies by RAM size and macOS version. Practically, a 36 GB MacBook Pro can load models that would require a 24 GB discrete GPU on a PC. Local-LLM tools show you the usable limit directly — see the LM Studio guide and the Ollama guide for how each reports it. The trade-off is memory bandwidth, which varies widely between base, Pro, Max, and Ultra chips and directly affects tokens per second.
How to Check VRAM on Linux
NVIDIA: nvidia-smi
Installed with the proprietary driver. Run nvidia-smi for a snapshot, or refresh every second while a model loads:
nvidia-smi -l 1AMD: rocm-smi or sysfs
With the ROCm stack installed:
rocm-smi --showmeminfo vramWithout ROCm, the amdgpu kernel driver exposes the total directly (in bytes; your card may be card1 — check ls /sys/class/drm/):
cat /sys/class/drm/card0/device/mem_info_vram_totalAny GPU: glxinfo
Install mesa-utils (Debian/Ubuntu) or glx-utils (Fedora), then:
glxinfo -B | grep -i memoryMesa drivers print a line like Video memory: 8192MB; the exact label varies by driver.
lspci — identifies the card, not the VRAM
lspci | grep -iE 'vga|3d'sudo lspci -v -s <slot> then shows the card’s memory apertures, but be careful: those are PCI BAR sizes, which only match total VRAM when Resizable BAR is enabled. Use lspci to identify the exact GPU model, then get the memory figure from the tools above or the card’s spec sheet.
Total vs. Free vs. Shared: What the Numbers Mean
| Number | Where you see it | What it means for LLMs |
|---|---|---|
| Dedicated VRAM (total) | Task Manager “Dedicated GPU memory”; nvidia-smi total | Physical memory on the card. This is the hard budget for model weights. |
| Free VRAM | nvidia-smi memory.free | What is available right now. The desktop, browser tabs, and other apps typically hold 0.5–2 GB. |
| Shared GPU memory | Task Manager “Shared GPU memory” | System RAM (up to about half of it) the GPU can spill into over PCIe. Far slower than VRAM — do not count it when sizing a model. |
| Unified memory (Apple Silicon) | About This Mac “Memory” | One CPU+GPU pool. The GPU can use most of it, so it behaves like a large VRAM pool with chip-dependent bandwidth. |
When loading a model, what matters is free VRAM, not total. A 12 GB card with 1.5 GB already consumed by the compositor and a browser has about 10.5 GB to work with — enough to make a model that “should fit” fail to load. On NVIDIA, the per-process list at the bottom of nvidia-smi output tells you exactly what to close.
How Much VRAM Do You Actually Need?
The question behind the question. At 4-bit quantization — the standard for local inference — model weights take roughly 0.5–0.6 GB per billion parameters, plus a KV cache that grows with context length. Rough guide:
| Model size | 4-bit weights (approx.) | VRAM to run comfortably |
|---|---|---|
| 7–8B | 4–5 GB | 6–8 GB |
| 12–14B | 7–9 GB | 10–12 GB |
| 24–32B | 13–19 GB | 16–24 GB |
| 70B | 36–42 GB | 48 GB, or split across two GPUs |
The “comfortable” column assumes moderate context (4k–8k tokens); very long contexts add multiple gigabytes of KV cache on top. For per-model figures, see the VRAM requirements table for every major LLM or browse the models database. For your exact combination of model, quantization, and context length, use the VRAM calculator. If your card comes up short, the best GPUs for local LLMs guide ranks upgrade options by VRAM per dollar — and if a new GPU does not pencil out, the self-hosting vs. API break-even calculator shows when paying per token is cheaper.
Frequently Asked Questions
Can I increase the VRAM on my graphics card?
Not on a discrete GPU — the memory chips are soldered to the board, so the only upgrade is a different card. Integrated GPUs are the exception: some BIOS/UEFI setups let you change the RAM allocation (often labelled “UMA Frame Buffer Size” or similar). Windows’ shared GPU memory is not extra VRAM and does not change what fits at full speed.
Why does Windows report more GPU memory than my card has?
Task Manager’s “GPU Memory” line is dedicated VRAM plus shared system memory, and dxdiag can inflate its figure the same way. The physical VRAM is the “Dedicated GPU memory” value, or the total that nvidia-smi prints. An 8 GB card on a 32 GB PC will often show 24 GB of “GPU Memory” — only 8 GB of it is real VRAM.
Does shared GPU memory help run larger LLMs?
Not meaningfully. Runtimes like llama.cpp and Ollama can deliberately offload layers to system RAM, which lets an oversized model run at all — but generation speed typically drops from tens of tokens per second to low single digits. Size your model to dedicated VRAM and treat RAM offload as a fallback, not a plan.
How much VRAM does a 7B or 8B model need?
About 4–5 GB for the weights at 4-bit, so a 6–8 GB card runs one comfortably with room for context. An 8 GB GPU handles the 7–8B class well; 12 GB opens up 12–14B models. The best local models for Ollama guide maps popular models to VRAM tiers.
Is unified memory on a Mac the same as VRAM?
For running LLMs, close enough: the GPU addresses one shared pool, up to a system-set limit of roughly two-thirds to three-quarters of total RAM. That means a 32 GB Mac can run models a 12 GB PC GPU cannot. The difference shows up in memory bandwidth, which varies several-fold between base and Max/Ultra chips and sets your tokens-per-second ceiling.
How do I monitor VRAM usage while a model is running?
On NVIDIA (any OS), run nvidia-smi -l 1 for a once-per-second refresh. On Windows, the Task Manager GPU pane updates live. On AMD Linux, use watch -n 1 rocm-smi --showmeminfo vram; on a Mac, model weights show up as ordinary process memory in Activity Monitor’s Memory tab, since the pool is unified.
