Overview
To check how much VRAM you have, run nvidia-smi on an NVIDIA GPU or amd-smi static -v on an AMD GPU, and read the memory size the tool prints; a Mac has no separate VRAM, so use its total unified memory instead. The two commands below come from the vendors’ own documentation, read on 2026-10-13. Once you have the number, vram-requirements shows what fits, and the formulas are in methodology.
How do I check how much VRAM I have?
| Your GPU | Command | What the vendor doc says |
|---|---|---|
| NVIDIA (Windows or Linux) | nvidia-smi -q | The nvidia-smi documentation lists “FB Memory Usage” with Total (“Total size of FB memory”), Used and Free. It also notes “Reported total memory can be affected by ECC state.” |
| AMD | amd-smi static -v | The AMD SMI CLI page describes static -v (--vram) as “All vram information”. The page’s sample output for plain amd-smi static has a VRAM section with type, vendor, size, bit width and maximum bandwidth; it has no separate sample for -v. |
| Apple Silicon | No separate VRAM | Memory is unified. See what fits my RAM. |
Windows Task Manager may also show GPU memory, but the Microsoft support pages tried on 2026-10-13 returned HTTP 404, so whether and where it shows is unverified here and no steps are given.
If you do not know which card you have, the model name tells you the memory. For example the RTX 5090 page lists 32 GB GDDR7. Other cards are on gpu-vram-list.
What is VRAM?
VRAM is the memory on a graphics card. NVIDIA’s tool calls the same thing “FB memory” (frame buffer). It is separate from system RAM; see ram-vs-vram.
How do I check VRAM usage while a model is running?
Use the same tools while the model is loaded. nvidia-smi reports Used and Free under FB Memory Usage. For AMD, the AMD SMI page lists amd-smi metric -m (“Memory usage per block”) and amd-smi monitor --vram-usage for live usage in MB. AMD says AMD SMI is “intended to replace and deprecate the existing rocm_smi CLI tool”, and the rocm_smi_lib page recommends moving to AMD SMI.
If you use Ollama, ollama ps shows whether the model landed on the GPU. Per the Ollama FAQ, 100% GPU means the model “was loaded entirely into the GPU”, 100% CPU means it “was loaded entirely in system memory”, and 48%/52% CPU/GPU means it was split between the two. A split suggests the model did not fit entirely in VRAM (our inference; the FAQ only describes the split).
Why is VRAM important for an LLM?
The weights and the KV cache must fit in VRAM to run fully on the GPU. When they do not, the runtime offloads to system RAM or fails with cuda-out-of-memory-llm. Speed in tokens per second is unknown here: no vendor page read for this article publishes a figure.
Will an LLM fit in the VRAM I have?
Use the site formula from methodology: weights_GB = params x bits / 8 / 1e9, plus KV cache, plus an assumed 1 to 2 GB overhead (an assumption, not a measurement). Worked example from vram-requirements, using the Qwen2.5-7B model card and llama.cpp’s Q4_K_M value of 4.8944 bits per weight: 7.61e9 x 4.8944 / 8 / 1e9 = 4.6558 GB of weights, plus 0.2349 GB of KV cache at 4,096 context, plus 1 to 2 GB, gives 5.8907 to 6.8907 GB. That is under 8 GB at 4,096 context, but longer context grows the KV term; see 8gb-vram-llm. For other sizes see vram-requirements, 16gb-vram-llm and 24gb-vs-16gb-vs-12gb-vram; a calculator is planned at vram-calculator.
Questions
How do I check how much VRAM I have?
Run nvidia-smi -q on NVIDIA and read Total under FB Memory Usage, or run amd-smi static -v on AMD and read the VRAM size. On a Mac there is no separate VRAM, only unified memory.
What is VRAM?
VRAM is the dedicated memory on a graphics card, separate from system RAM. NVIDIA’s nvidia-smi calls it FB memory.
Why is VRAM important for an LLM?
The model weights and KV cache must fit in it for the model to run fully on the GPU. If they do not, the runtime offloads to system RAM or fails.
How do I check VRAM usage while a model is running?
Run nvidia-smi and read Used and Free, or on AMD run amd-smi monitor --vram-usage. With Ollama, ollama ps shows the CPU/GPU split in its Processor column.
Will an LLM fit in the VRAM I have?
Add weights (params x bits / 8 / 1e9), KV cache, and an assumed 1 to 2 GB of overhead, and compare the total with your card. A 7B model at Q4_K_M and 4,096 context comes to 5.8907 to 6.8907 GB by that method.
Does the Windows Task Manager show VRAM?
Unknown here. The Microsoft support pages could not be fetched on 2026-10-13, so this page does not give Task Manager steps.
Sources read on 2026-10-13
- nvidia-smi documentation: FB Memory Usage Total/Used/Free, ECC note. The
memory.totalquery field name was not found in the text read, so it is not used. - AMD SMI CLI:
static -v(“All vram information”), samplestaticVRAM fields,metric -m,monitor --vram-usage, replacement of rocm_smi. - rocm_smi_lib: recommendation to move to AMD SMI.
- NVIDIA RTX 5090: 32 GB GDDR7.
- Ollama FAQ:
ollama psProcessor column. - Worked example figures are from vram-requirements (llama.cpp quantize README and Qwen2.5-7B card, read there).
- Microsoft Task Manager support pages: HTTP 404, not used.
Related
- vram-requirements sizes by model.
- vram-calculator and gpu-vram-list size by card.
- ram-vs-vram explains the two memory types.
- 24gb-vs-16gb-vs-12gb-vram, 8gb-vram-llm, 16gb-vram-llm by tier.
- cuda-out-of-memory-llm covers failures.
- methodology is the citation rule.
- Mac: what fits my RAM.