Neural Rendering can keep the graphics card under a limit, and every install now runs the model on every 2nd frame. The limit is a checkbox under Auto quality (off by default): when the card sits at the limit, the model gets lighter. It was added after a report from an RTX 4090 at 99 % load, where limiting the card to 95 % cleared the problem. Tested offline (auto quality) and with the sensor on one RTX 4070 Ti SUPER; the checkbox itself has not been tried in a game yet.
- Neural Rendering: "Keep the graphics card under a limit" (under Auto, off by default, 95 % unless moved). A reporter with an RTX 4090 at 99 % load in WoW (their own overlay) had the problem go away when they limited the card to 95 %; the log showed the model waiting about 4 ms behind the game's work. With the box ticked, auto quality reads the card's load once a second (NVML, NVIDIA cards only;
gpu_load.cpp) and, when it stays at or above the limit for 3 s, treats it like slow game frames: the budget shrinks to 60 %, the resolution comes down, it does not go back up for a minute, and at the lowest resolution the model runs on every 2nd or 3rd frame. It lets go after 10 s at 8 points under the limit. It only takes the model's share off the card: a game that fills the card alone still needs a frame rate limit.nr_autotestcovers it,nr_gpuloadtestreads the sensor. - Neural Rendering: every install gets "Run the model every 2nd frame", including one that ran 0.9.28. 0.9.29 fixed a config that never had the setting, but 0.9.28 itself had loaded such a config as "every frame" and saved that back, so those kept it. A saved 1 from a file written before the defaults' version (a new key,
defaultsVersion) now becomes 2, once; a 1 chosen after that stays.nr_settingstestcovers it.
Zip LSAddonManager-0.9.30-x64.zip, SHA-256 c9665192790e72c312423cb4181cae6a9010390b78d820bf9cafa6729b7e8dba.