VRAM + RAM Offload Calculator for Local LLMs
Use this when a model is too big for VRAM but system RAM is available. It estimates whether the model fits, how many transformer layers may fit on the GPU, and how much weight memory spills to RAM. Competing calculators also model CPU/RAM offload; this version keeps its assumptions visible and accepts custom architecture inputs.
VRAM + RAM Offload Calculator
Enter the VRAM and system RAM you actually have, and see whether a model fits, how many layers may stay on the GPU, and how severe the likely offload penalty is.
When a model is too big for VRAM, llama.cpp can keep some layers on the GPU and the rest in system RAM. This works out the split for you.
Fits, but heavily offloaded — most layers run on the CPU/RAM, so generation will be slow. A bigger GPU or a smaller quant would help most.
--n-gpu-layers 26 of 80| Model weights (70B at Q4) | 35GB |
| KV cache (8K context) | 2.7GB |
| Runtime overhead | 1.5GB |
| Total memory needed | 39.2GB |
| You have (VRAM + RAM) | 16GB + 64GB |
These are planning estimates: weights are parameters × bytes-per-parameter; FP16 KV cache uses the model's layers, KV heads, head dimension, and context; and speed is a qualitative tier. Quantized cache, sliding-window attention, runtime behavior, and unequal layer sizes can change the result. For the full memory rules see the VRAM requirements guide, or size a card with the VRAM calculator and the GPU Value Finder.
What you enter
- Your VRAM
- Your system RAM
- Model or custom architecture
- Quantization
- Context length
What it returns
- Estimated full-GPU, offload, or no-fit status
- Approximate --n-gpu-layers value
- Estimated weights spilled to RAM
- A memory breakdown (weights, FP16 KV cache, overhead)
- A qualitative speed tier
Unique value
The useful addition is architecture-aware FP16 KV-cache math and custom parameter/layer/KV-head/head-dimension input alongside the VRAM/RAM layer estimate. It is a planning aid, not an exact runtime prediction or an uncontested category.