Notice. Not affiliated with NVIDIA; NVIDIA, DLSS and the NVIDIA logo are NVIDIA Corporation's trademarks. The bundled NVIDIA runtimes are NVIDIA's property, included unmodified as received, research/educational use only, no warranty, use at your own risk.
Warning
NeuralScreen requires the latest NVIDIA driver. Operation with older drivers or unsupported/non-standard configurations is not guaranteed.
The picture no longer stutters under an open panel, a GPU driver reset is handled more carefully on RTX 20-40, NR starts the way you left it, and the NR preset switch appears only where it does something.
Fixes
- An open panel no longer stutters the picture under it (#166). The panel's live readings changed with every frame, and each change redrew the overlay layer, which costs the compositor a frame of timing. They now change twice a second while the panel is open. Bench, a 71.93 fps source on a 144 Hz screen with the panel open: 7.4% of source frames lost and 87.1% evenly spaced before, 2.0% and 98.1% now (panel closed: 0% and 99.5%).
- After a GPU driver reset, NeuralScreen waits for the driver (#181). On an RTX 3090 the driver reset the GPU under NR, NeuralScreen restarted the worker within two seconds while the driver was still recovering, and the PC froze. It now waits until the driver answers before restarting; on RTX 20-40 (where NVIDIA's runtime runs through an architecture patch) a second reset in one session leaves NR off, and Num1 turns it back on. The reset itself happens inside NVIDIA's runtime on those cards.
- The NR preset switch shows only on a runtime with two models (#179). The bundled NR runtime (310.8) has one model, so A and B ran the same. Fine tuning now reads the runtime's version: with 310.9.16 or newer in
native\libraries\the Auto / A / B switch is there, otherwise one line says "one model (310.8.0.0)".
New
- NR starts the way you left it (#166). Num1 and the NR tile save the state; Frame Generation already did.
Tests
Full suite: 331 pass, 2 fail, 6 skipped; GUI-E2E 14/14. The two, test_fg_fresh_frames and test_gpu_recorder, pass on their own (the recorder's check is native and unchanged since 2.6.0). The bench run above is with RTSS and PresentMon closed.