About the VLA & Robotics Edge VRAM Calculator
The VLA & Robotics Edge VRAM Calculator (vla-vram-calc.physicaiportal.com) is a free browser tool that helps robot developers estimate the unified memory and memory bandwidth a Vision-Language-Action (VLA) model needs before deploying it on an onboard computer such as NVIDIA Jetson Orin or Thor. It is available in Korean, English, Japanese, Chinese (Simplified and Traditional), German, French and Spanish.
1. What the calculator estimates
- Model weights: parameter count × bytes per parameter for the chosen precision (FP16 2.0, FP8 1.0, W8A16 about 1.05, W4A16 with an FP16 action head about 0.62, NVFP4 about 0.56).
- Vision-token KV cache and activations: an empirical estimate proportional to cameras × tokens per camera × history frames and to model size.
- 3 GB system reserve: ROS 2 communication buffers, camera buffers, CUDA/TensorRT runtime and OS headroom.
- Bandwidth and inference rate: inference rate = control rate ÷ actions executed per chunk; required bandwidth = memory read per inference × inference rate ÷ bandwidth utilization.
- Hardware data: memory, bandwidth and power from NVIDIA's official Jetson Orin and Thor product pages. Model parameter counts come from each model's paper, official repository or model card.
2. Disclaimer
Results are estimates to support early design and hardware selection. For real robot deployments, and especially for safety-critical control loops, always confirm against vendor datasheets and memory and latency measurements on the target device. See the disclaimer and terms for details.
3. Corrections and contact
If you find a changed hardware specification, a better formula, a translation error or a bug, email [email protected]. Coefficients and reference data are updated after review.