Nvidia Greenboost — On Our Radar

Status: 🟡 Under observation
Why we're talking about it: This unofficial community hack is gaining traction for using CUDA’s DMA-BUF path to map system RAM and NVMe storage as extended VRAM.

What we know:
- The project uses the documented but niche cudaImportExternalMemory path to load models exceeding physical VRAM capacity (source: GitLab).
- User reports from March 14, 2026, confirm a 12GB RTX 5070 successfully loaded a 32GB GLM-4.7 model (source: User Reports).
- Performance is severely bottlenecked by PCIe 4.0/5.0 bandwidth, resulting in 2-5 tokens per second for models over 30GB (source: GitLab).
- It operates at the CUDA allocation layer and functions alongside official Nvidia drivers rather than replacing them (source: Project Readme).
- System instability is a high risk, as the Linux kernel may trigger the OOM killer during heavy memory mapping (source: HN).

The unknowns:
- Performance scaling for DDR5 remains undocumented, as current benchmarks focus primarily on DDR4-3600 (source: GitLab).
- Support status for Windows or WSL2 is currently unconfirmed or missing from the documentation.

This article will be updated when we have more data. Until then, proceed with caution.

Ship clean code,
Marcus.