I'm using a GeForce RTX 3090 ASUS ROG Strix OC Edition (previously used for mining) for LoRA training, which is pushing the card to its limits. My setup:
- Windows 10
- 64GB RAM
- ROG Crosshair VIII Hero MOBO
Issue: Memory temperatures spiking past 100°C during extended LoRA training sessions.
Current solution:
- Using MSI Afterburner with a voltage curve based on this video: https://www.youtube.com/watch?v=SIlXT32fOMk
- Results: Memory temps now between 94-96°C, GPU and Hot Spot temps down by ~3°C
While this is an improvement, I'm aiming for even lower memory temps for safer long-term use. I'm willing to trade some performance for better thermal management.
Questions:
- Does anyone have experience with optimizing RTX 3090 settings specifically for ML workloads like LoRA training?
- Are there better MSI Afterburner curves or settings you'd recommend for my use case?
- Any additional tips for managing VRAM temperatures on the 3090 during intensive, extended workloads?
Thanks in advance for any insights!