NVidia GreenBoost kernel modules opensourced
New kernel module lets AI models run on GPUs with 2-3x less VRAM, no code changes needed.
NVIDIA has taken a significant step towards democratizing access to large-scale AI by open-sourcing its GreenBoost technology. This release consists of a Linux kernel module paired with a CUDA userspace shim, creating a transparent memory extension layer. The system works by intercepting calls to GPU memory management functions, allowing unused portions of AI models or computational graphs to be automatically swapped out to the system's much larger DDR4 RAM or even high-speed NVMe storage. Crucially, this happens without any modification to the underlying AI inference software, meaning existing applications like LLM servers, ComfyUI workflows, or Wan2GP tools can immediately leverage larger models.
The practical impact is substantial for developers and researchers working with constrained hardware. A user with an 8GB GPU could effectively run models that previously required 16GB or 24GB of VRAM, albeit with a performance trade-off due to the latency of moving data between GPU memory, system RAM, and storage. The technology is not limited to language models; any CUDA-based application that performs VRAM detection and allocation—including emerging AI desktop agents and creative tools—can benefit. This opens the door for more rapid prototyping and lower-cost experimentation with state-of-the-art models that were previously gatekept by expensive hardware requirements.
- Transparently extends GPU VRAM using system DDR4 RAM and NVMe storage as a swap space.
- Requires zero modifications to existing AI inference software like LLM servers or ComfyUI.
- Enables running models that require 2-3x the physical VRAM available on the GPU hardware.
Why It Matters
Lowers the hardware barrier for running cutting-edge AI, enabling broader access to large models for developers and researchers.