NOTICE: New (~5/30?) NVidia drivers change memory management

[cgessai]

PREVIOUS BEHAVIOR: If more VRAM was requested than was available, error would result

NEW BEHAVIOR: If more VRAM is requested than available, drivers seamlessly and automatically dip into Shared GPU Memory (i.e. RAM).

More information, but related to Automatic1111 (stable diffusion). Just be aware going forward things may behave non-intuitively based on previous performance. After upgrading I chose to roll back.

[LostRuins]

Good to know. I guess people may be surprised when offloading more layers results in slower performance. Not sure how to handle this.

[brknsoul]

If this is still going, iirc the last driver version before nvidia went in this direction was 531.

[pacmanincarnate]

is there any idea if pushing into shared memory is slower than running off of RAM? If I'm splitting a model between VRAM and RAM already, is there an efficiency loss in increasing my GPU layers and pushing more into shared vs just RAM?

[berkut1]

The Nvidia community manager said that they will fix that in the next driver update. vladmandic/sdnext#1285 (reply in thread)