Intentionally reducing VRAM available

[porina]

There's a lot of discussion on how much vram is "needed" looking forwards, but I thought it might be interesting to see if we can test it going backwards. Of course, we can't magically add more VRAM to a card, but is there a way we can effectively reduce the VRAM? There was something on the news not that long ago about someone creating a ramdisk in VRAM so that might be one way to do it. The only requirement is that the VRAM is locked out of graphics usage, and that whatever is locking it doesn't consume other GPU resources to affect gaming. So a compute workload is not suitable for example.

 

Then we could go to the GPUs. The 1080Ti and 2080Ti both have 11GB of VRAM. The 3080 is on 10GB. AMD 6800 cards have 16GB, and many older mid-high cards are typically around 8GB. So for starters, limit those above 8GB to 8GB, how does that affect gaming performance? Reduce it to 4GB, what is the impact there? Of course, this will vary a LOT between different games, and different settings. It is no small job, but I think it would be an interesting exercise to understand the impact.

 

For 30 series and 6800 series, it would also be interesting to see if the higher PCIe 4.0 bandwidth would help if there are situations where lower available VRAM does cause a performance impact.

[Skiiwee29]

This would functionally be accomplished by using a 3070 vs 2080ti... 8gb vs 11gb... since they effectively are basically identical in Rasterization. 

[WereCat]

Just now, Skiiwee29 said:

This would functionally be accomplished by using a 3070 vs 2080ti... 8gb vs 11gb... since they effectively are basically identical in Rasterization. 

not really a good way to compare for accuracy reasons because of different architectures, there may be situations where 2080ti will absolutely beat the 3070 and vice versa,

 

@porina

https://github.com/prsyahmi/GpuRamDrive

 

Just use the GPU VRAM disk and fill it up completely so that the VRAM stays allocated and then run the workload? Wouldn't that work?

[porina]

3 minutes ago, Skiiwee29 said:

This would functionally be accomplished by using a 3070 vs 2080ti... 8gb vs 11gb... since they effectively are basically identical in Rasterization. 

Too many variables. The architecture and core configuration are very different between them. The best result would be using specific cards and effectively reducing the VRAM on them to see the impact, and if this project were to go ahead, I'd like to go below 8GB also.