PSA: Nvidia Driver 535.98

[scrozier]

Hello!

Can't say for certain but I feel like this driver has improved the UI snappy-ness of premiere by 2x especially when doing things like repositioning in the program monitor. When clicking on "motion" and directly manipulating things like mask points and position in the program monitor, its really smooth now when last week it was a lag-fest.  

https://www.nvidia.com/download/driverResults.aspx/205035/en-us/ 

I'm running 23.3 (23.4 seems like it has some bad bugs I'm trying to avoid for now)

[RjL190365]

I tried this new driver on both my main Intel i7-12700K system that's equipped with a GeForce RTX 4070 Ti GPU and on my secondary AMD Ryzen 7 3700X system that's equipped with a GeForce GTX 1650 SUPER GPU. For some reason the new driver appears to penalize certain older and/or lower-end GPUs' CUDA performance while keeping the performance of the newer higher-end GPUs relatively unchanged. In the PugetBench for Premiere Pro 0.98 export tests, the GPU effects score using the Standard preset on my GTX 1650 SUPER plummeted from 24+ with 531.61 all the way down to only 13-ish with 535.98. The RTX 4070 Ti's performance was very much equal - between 68 and 69 - with either driver version.

This makes me want to upgrade the GPU of my secondary rig to an RTX 3060 or an RTX 4060 in the near future. The new driver version seems to dislike GPUs with a lower number of CUDA cores; after all, the RTX 1650 SUPER has only 1280 CUDA cores while the RTX 4070 Ti has 7680.

By the way, the Production Branch of the Nvidia Workstation GPU driver (for Quadro and non-GeForce-branded workstation GPUs such as the T1000 and the RTX A4000) has also been updated to 535.98 as well. This came after complaints of compatibility issues with the 528.89 and 528.95 drivers.

[RjL190365]

An update to my testing with the GTX 1650 SUPER with driver version 535.98:

It turned out that my Extended preset tests for PugetBench consumed so much VRAM that the renderer got ungracefully slammed into the software-only mode in the middle of my testing, which skewed both my Standard and Extended scores noticeably downwards, especially in the GPU Effects part of the scoring.

I discovered this after running only the Standard preset in PugetBench, which did produce scores that were where they should have been for this GPU. And GPUs with only 4 GB of VRAM cannot properly run the Extended preset tests, as evidenced by the significantly lower GPU Effects score in the Extended preset than in the Standard preset when the GPU was on driver version 531.61.

So it wasn't the CUDA shortage after all, but instead the VRAM got depleted during the Extended preset testing.