Built openNR's compat-fp16 DLSS-NR 310.9.16 locally — bit-identical hash, NR cost 6.07 ms → ~3.5 ms on an RTX 4070 (Dead as Disco, OptiScaler F5)

[On-100K-Lisse]

Hi all — sharing a data point for anyone experimenting with DLSS Ray Reconstruction (DLSS-NR) DLLs through OptiScaler on RTX 40-series.

What I did

Using ShortFuse's open-source openNR toolchain, I built the compat-fp16 variant of nvngx_dlssnr.dll 310.9.16.0 locally (the build keeps NVIDIA's network weights and adds GPU code for RTX 20/30/40/50). Toolchain: Visual Studio 2026, CMake 4.4.4, CUDA Toolkit 13.4.92 (nvcc V13.4.92 — the pinned version), 45 cubins compiled via ptxas.

One build gotcha worth knowing: build-tools.cmd and CMakePresets.json hardcode the Visual Studio 17 2022 generator. With only VS 2026 (v18) installed, CMake fails with "could not find any instance of Visual Studio". Replacing the generator with Visual Studio 18 2026 in both files (and deleting the stale build/ cache) fixes it — and since the toolchain self-verifies against the reference hash, a wrong generator would have been caught anyway.

Verification

The build is self-checking, and mine passed:

Exact reference match: CDFA6C1EC784867182164C7836F324F087B5C9FE46F54999FEABF3ED

Bit-identical to ShortFuse's reference build — no trust-me-bro, just a hash.

In-game results (OptiScaler F5 fork v0.1.32, Dead as Disco, RTX 4070 12 GB, 2560×1440, FG off, same scene/settings, one variable at a time)

DLL	NR status	NR GPU time	FPS (capture)
310.8.SF.0 (community build)	Running	6.07 ms	~68
310.9.16 compat-fp16 (openNR build)	Running (NGX, preset B)	3.43–3.65 ms	82.8

So roughly 40% lower NR cost than the 310.8 build on an Ada card — in line with what others have reported on RTX 4060/4080-class hardware.

Also confirmed through the RenoDX/ReShade path: NR INJECTED, 4,856 successful NR frames, NGX 0x00000001. Amusing detail: the RenoDX panel labelled the DLL "310.8.0" while its hash proved it was the 310.9.16 compat build — labels lie, hashes don't. The same DLL also runs in Crime Scene Cleaner via the feeder path (17,801 NR frames).

Limits, honestly

• Measurements are from one title with spot captures, not a controlled benchmark suite.
• Preset A (main model) not yet tested with the compat-fp16 build — on the list.
• This is a reproducible-build + measurement report, not a DLL distribution: openNR builds the DLL from your own extracted files; I only share the documentation.

Full write-up (build log, test matrix, failure analysis of the stock DLL's architecture check, per-game results): https://github.com/On-100K-Lisse/travail-avec-ia/blob/ai/kai/dlss5/docs/DLSS5_310.9.16_RTX4070.md — heads-up, the documentation is in French; I'm happy to translate any section here if there's interest.

Happy to answer questions about the build or the measurement protocol — and if anyone reproduces the numbers on other Ada/Ampere cards, I'd love to compare.