Without changing a single line of code, official DLSS runs on the RX 7700 XT: measured at 72 FPS. AMD's graphics card has now run NVIDIA's original DLSS. A developer with the online handle countervolts released a project called d4r, short for DLSS 4 Radeon, which brings NVIDIA's official DLSS library nvngx_dlss.dll intact

💡 What You Will Learn

Without changing a single line of code, official DLSS runs on the RX 7700 XT: measured at 72 FPS. AMD's graphics card has now run NVIDIA's original DLSS. A developer with the online handle countervolt

📜 Table of Contents

• Not a Single Line of the Library Changed, CUDA Calls Intercepted Mid-Route
• The Real Hard Part: Running DLSS 4 Directly Produces Wrong Results
• The Cost Is All in the Numbers
• Want to Try It? Clear Three Hurdles First

No Code Changes, Official DLSS Runs on RX 7700 XT: 72 FPS Measured

An AMD graphics card has now run NVIDIA's original DLSS. A developer going by the handle countervolts released a project called d4r, short for DLSS 4 Radeon, which ports NVIDIA's official DLSS library nvngx_dlss.dll unmodified onto the RX 7700 XT, running under Linux via Proton. In a measured test of Silent Hill Townfall at 2560×1440 Quality mode, DLSS 3 mode hit 72.4 FPS, with all three DLSS models rendering correctly.

Not a Single Line of the Library Changed, CUDA Calls Intercepted Mid-Route

First, why this is remarkable. DLSS's core is a dynamic link library packed with CUDA code, which by nature only recognizes NVIDIA graphics cards — install it on an AMD card and it won't recognize it either. d4r's approach is to build a bridge in between: when a game calls DLSS, Wine hands these CUDA calls over to ZLUDA. ZLUDA is an open-source CUDA compatibility layer with 14.9k stars on GitHub, whose job is to translate NVIDIA's PTX instructions (an intermediate instruction format NVIDIA designed for its own GPUs) into instructions AMD cards can execute, running on ROCm/HIP underneath. Once the translation is done, the DLSS library itself can't even tell it's no longer on an NVIDIA card. The whole pipeline has an absurd number of steps, and VideoCardz's assessment is that achieving this level of performance after so many layers of translation is crazy in itself.

The Real Hard Part: Running DLSS 4 Directly Produces Wrong Results

Being able to translate doesn't mean being able to compute correctly. d4r supports DLSS 3's CNN model, DLSS 4 transformer, and DLSS 4.5 transformer, but DLSS 4 running directly through ZLUDA produces incorrect output, and upon detecting the anomaly, DLSS simply discards the entire network output. The developer had no choice but to rewrite the most expensive kernel, replacing it with a hand-written custom version targeting the RDNA3 architecture, which is what saved DLSS 4. This custom kernel is currently fully optimized only for gfx1101 (Navi 32 core), covering four cards, with the developer having only tested the RX 7700 XT. Support across cores is as follows:

Reading this table makes it clear that, strictly speaking, this is a Navi 32-exclusive DLSS patch, not something the entire RDNA3 lineup can use.

The Cost Is All in the Numbers

Measured data for the RX 7700 XT at 2K resolution: Quality mode, DLSS 3 CNN's Preset E at 72.4 FPS, DLSS 4's Preset K at 69.4 FPS, and DLSS 4.5's Preset M at only 51.5 FPS. DLSS 4's overhead is essentially fixed at 2.8ms, because its network computes based on output resolution; DLSS 4.5 eats 7.5ms of GPU time per frame, only overtaking at Ultra Performance mode, 94.5 FPS versus 88.9 FPS. What stings more is that across all modes the developer tested, AMD's own FSR 4 is faster than this DLSS setup. After all that effort to bring NVIDIA's library over, the fastest option is still Su's own solution.

Want to Try It? Clear Three Hurdles First

First, it currently only runs under Linux and Proton — Windows players can't use it. Second, DLSS 5 is explicitly unsupported; only the three models DLSS 3, DLSS 4, and DLSS 4.5 work. Third, all data comes from a single developer's own testing, with no third-party reproduction yet — VideoCardz is also just relaying his submission. According to VideoCardz's speculation, someone may later port it into an OptiScaler plugin, and that's when it'll become a version ordinary players can install directly. The modding scene has been busy these past few months. Some have unlocked DLSS 5 and frame generation for older RTX 20/30/40 series cards, others have gotten DLSS 5 running on integrated graphics or even in a browser, but all of those efforts target NVIDIA's own hardware. This time is different — d4r goes after the CUDA compatibility layer, prying open a crack in NVIDIA's most core software moat. The closed-source AI libraries that have sat in the CUDA ecosystem for over a decade now theoretically all have a path onto A cards. If you've got a 7800 XT in hand, would you go through the trouble of setting up a Linux environment for a set of DLSS?