Skip to content

GraphCast-Lite: Running 0.25° (37 levels) FP32 model on 8GB VRAM consumer GPUs #218

Description

@yxdragon

Hi GraphCast Team,
Thank you for your groundbreaking work!
I'd like to share a community project, GraphCast-Lite, designed to run GraphCast inference on consumer hardware by heavily optimizing memory and speed.
Key Highlights:
8GB VRAM Support: Runs the high-resolution 0.25° model (37 levels) in full FP32 precision on consumer GPUs (e.g., RTX 2080 Ti) using a Static VRAM Pool and Dual-Chunking Engine.
High Performance: Forecasts 10 days (40 steps) in ~15s on RTX 4090 (BF16) and ~40s on RTX 2080 Ti (FP16) via TensorRT.
Lightweight: Replaces heavy dependencies with a pure PyTorch/TensorRT inference pipeline.
Repo: <VectorElectron/graphcast-lite>
Hope this helps more researchers experiment with GraphCast without needing enterprise GPUs!
Best regards,

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions