DLSS 5 on RTX 4070 Ti: OptiScaler’s NR Before SR Boosts FPS in Elden Ring Test

What Nvidia actually supports​

Nvidia's announcement says DLSS 5 debuted in NBA 2K27 for GeForce RTX 50 Series GPUs and laptops, plus GeForce NOW. Nvidia's research page says the feature runs locally as a rendering stage within existing game pipelines on GeForce RTX 50 Series GPUs.

An RTX 4070 Ti is Ada Lovelace, so it falls outside that official support. XDA's author gets around this with OptiScaler, a spoofed GPU identity and the neural rendering DLL files that have circulated online. None of that is a supported Nvidia path.

Nvidia describes DLSS 5 differently from earlier DLSS features. In its description, DLSS 5 is the first DLSS technology to generate the final displayed appearance, rather than reconstruct a higher-cost reference output. Super Resolution reconstructs a higher-resolution image, and neural rendering (NR) changes how the image looks. That difference explains why the order of the two stages matters.

The setting: NR before SR​

By default, the modded neural pass runs on an image that Super Resolution has already upscaled. At 1440p or 4K, that means the model processes a full-size frame, and every extra pass repeats that work.

"NR before SR" reverses the order. Neural rendering runs on the lower internal-resolution frame, and Super Resolution upscales the result afterwards. The neural model gets fewer pixels to process. That is the whole trick.

This isn't unique to one build. Wccftech reported on a fork that lets the model process the lower-resolution input, such as a 1920x1080 frame in 4K Performance mode, before DLSS upscales it. Several community forks now offer pre-SR placement, and some add multi-pass options.

The reported Elden Ring numbers​

XDA's test system was an RTX 4070 Ti, a Ryzen 5 7600X and 32GB of DDR5. The game ran frame-unlocked at maximum settings with ray tracing. Here are the author's results:

I calculated the following from the author's table. These are not separate test results.

• Frame rate: Moving NR before SR raises the average by about 38% with one pass (71 to 98 FPS). It also gains about 38% with two passes (52 to 72) and about 45% with three (42 to 61).
• VRAM: In each pair, the before-SR run used less VRAM: 0.3GB less with one pass (10.1GB to 9.8GB) and three passes (11.8GB to 11.5GB), and 0.2GB less with two passes (10.7GB to 10.5GB).
• Baseline: Even the best before-SR result, 98 FPS with one pass, is well below the 147 FPS of DLSS 4.5 alone.

The author says the first pass recovers roughly half the performance lost under conventional NR. The table doesn't back that up as a general rule. It shows a gain in each pairing, but no fixed recovery rate.

What the numbers don't tell you​

The test leaves several things unspecified:

• the output resolution and DLSS preset behind the rendering numbers
• the driver version, game version and OptiScaler build
• which NR model or DLL was used
• how many runs were averaged, and whether frame times were measured
• any stability data

The author's account of crashes is anecdotal. He says three passes risk crashing on 12GB and that he stuck to two layers. The table shows 11.5GB used at three passes, which is close to the card's ceiling. Nothing here says 12GB is safe in other games, at other resolutions or in other scenes. Image quality is also the author's own judgment. He says the NR look remains clearly better than vanilla rendering even when NR runs before SR. No screenshots or measurements were provided.

A second XDA piece and other forks​

An earlier XDA article by the same outlet reports similar behaviour. It says pre-SR neural rendering isn't quite as pristine, but still offers an enormous visual jump over vanilla rendering. That is the author's opinion as well, but it hints at a real trade-off. Running the model on a lower-resolution frame may cost some fidelity.

Other community work points the same way. One fork's README says the pre-SR mode runs the model on the render-resolution image right before DLSS Super Resolution. That fork reports its own figures from a different scene and GPU. Its tables show a pre-SR "Quality" mode gaining about 36% over the stock approach. Those numbers are the fork author's, and they measure a different setup.

The gains therefore seem consistent in direction, though not in size. Treat the magnitude as hardware-, game- and build-dependent.

Risks and practical notes​

This is unofficial modding, and the forks say so. One fork's documentation calls it an experimental build. It warns that NR support, model behaviour, accepted presets and game compatibility may change between NVIDIA driver or model releases. The documentation also says Nvidia's proprietary model isn't distributed with the repository, and it advises keeping a backup of the original game files. Updates and file verification can also remove or replace proxy files.

Setup details vary by fork, so check the documentation for the build you use. One fork that documents the split pipeline uses the Insert key to open the overlay. It exposes an "Enabled" toggle and a "Run NR before SR" toggle, and it asks you to check the live status line and the OptiScaler log. I can't confirm that XDA used that build or those steps.

Be cautious about online play. Injecting DLLs into games can trigger anti-cheat systems, and I found nothing in the evidence on how particular games handle this. Single-player titles are the lower-risk place to experiment, but that is general advice and not a finding.

What this means for 12GB owners​

• The positive reading: a reordered pipeline can make an expensive neural effect far more usable on a mid-range, older GPU.
• The skeptical reading: DLSS 5 on Ada comes from third-party files and spoofing, so image quality, stability and compatibility are not guaranteed.
• The author's "software-defined" framing is interpretation. This experiment shows the effect can run on an RTX 4070 Ti. It does not show that every difference between GPU generations is a software restriction. Nvidia says its model uses RTX 50-series Tensor Cores, and performance on other hardware may differ.

In short, "NR before SR" is a real, documented pipeline option in experimental OptiScaler forks. In XDA's Elden Ring tests it improved frame rate and trimmed VRAM at every pass count. It doesn't remove the cost of neural rendering, and it doesn't make 12GB cards immune to VRAM limits.

Correction (November 7, 2026): An earlier version of this article said the before-SR run used 0.3GB less VRAM in each pair. That is correct for one and three neural-rendering passes, but with two passes the saving is 0.2GB (10.7GB to 10.5GB). The article's own table showed the correct figures; only the summary sentence was wrong.