How much VRAM do I have?
On Windows, press Ctrl+Shift+Esc, open Performance and select your GPU: "Dedicated GPU memory" is your VRAM. On a Mac, open the Apple menu and choose About This Mac. On Linux or a rented cloud GPU, run nvidia-smi. The steps for each are below, then how much memory the popular models need and the cheapest card to rent if yours is too small.
Windows 10 and 11
- Press Ctrl+Shift+Esc to open Task Manager and choose the Performance tab (select "More details" first if you see a small window).
- Click GPU 0 (or GPU 1 on a laptop with two graphics chips; the one with a model name such as NVIDIA GeForce RTX 4090 is the dedicated card).
- Read Dedicated GPU memory: that is your VRAM. "Shared GPU memory" is ordinary system RAM the GPU can borrow, much slower, and not what model memory guides mean.
A second way: press Windows+R, type dxdiag, and on the Display tab read "Display Memory (VRAM)".
macOS
Apple silicon (M1 and later): Apple menu, then About This Mac. The "Memory" figure is unified memory shared by the processor and the GPU; there is no separate VRAM number. The GPU can use a large share of it, so a Mac with more unified memory runs larger models, but not all of it is available to the GPU at once.
Intel Macs: Apple menu, About This Mac. The "Graphics" line names the GPU and its memory. For more detail, hold Option, choose System Information, then Graphics/Displays and read the VRAM line.
Linux and rented cloud GPUs
On an NVIDIA GPU, run:
nvidia-smi --query-gpu=name,memory.total,memory.used --format=csvIt prints one line per GPU with its total and used memory in MiB. Plain nvidia-smi shows the same in its "Memory-Usage" column, and nvidia-smi -L lists every GPU in the machine. On AMD GPUs with ROCm, run rocm-smi --showmeminfo vram.
On a rented machine, check before you start the job. Confirm that the GPU model and memory match what you rented (an "80 GB" listing should show an 80 GB card), and that you see as many GPUs as the node you are paying for. A GPU split into smaller slices (NVIDIA calls them MIG instances) shows only the slice's memory, not the whole card's.
How much VRAM do you need?
A model needs memory for its weights, for the key-value cache that grows with context, and for runtime headroom. These estimates use 4,096 tokens of context, one request at a time and 20% headroom; longer context or several users at once need more. Prices are the cheapest on-demand rental, checked 12 October 2026.
| Model | At 4-bit | Cheapest card for 4-bit | At 16-bit | Cheapest for 16-bit |
|---|---|---|---|---|
| Qwen2.5 7B | 4.9 GB | 16 GB RTX A4000 $0.10/hr | 18.6 GB | 24 GB RTX 3090 $0.20/hr |
| DeepSeek R1 Distill Llama 8Bsame architecture as Llama 3.1 8B | 5.5 GB | 16 GB RTX A4000 $0.10/hr | 19.9 GB | 24 GB RTX 3090 $0.20/hr |
| Qwen2.5 14B | 9.8 GB | 16 GB RTX A4000 $0.10/hr | 36.4 GB | 48 GB RTX A6000 $0.35/hr |
| Qwen3 32B | 20.9 GB | 24 GB RTX 3090 $0.20/hr | 79.9 GB | 80 GB A100 $1.09/hr |
| Mixtral 8x7B v0.1 | 28.7 GB | 48 GB RTX A6000 $0.35/hr | 113 GB | 192 GB MI300X $2.39/hr |
| DeepSeek R1 Distill Llama 70Bsame architecture as Llama 3.3 70B | 43.9 GB | 48 GB RTX A6000 $0.35/hr | 171 GB | 192 GB MI300X $2.39/hr |
| Qwen2.5 72B | 45.2 GB | 48 GB RTX A6000 $0.35/hr | 176 GB | 192 GB MI300X $2.39/hr |
Change the precision, context or batch for any model in the LLM VRAM calculator, or browse every model's memory page.
If your card is too small
Renting by the hour is the quick fix for a single job. The cheapest on-demand cards with enough memory today:
- 24 GB or more: 24 GB RTX 3090 from $0.20/hr (TensorDock)
- 48 GB or more: 48 GB RTX A6000 from $0.35/hr (Thunder Compute)
- 80 GB or more: 80 GB A100 from $1.09/hr (Thunder Compute)
Buying a card for a desk instead? Compare retail graphics card prices on ComponentPrices, then weigh the purchase against renting with the rent or buy calculator.