The DLSS 5 Back-Ports Show Where the Hardware Line Really Sits

Modders have pushed the leaked DLSS 5 runtime from Blackwell down to Ada, Ampere and now Turing. The frame rates say more about NVIDIA's likely support list than the ports do.

Saya4 minute readDrivers & Software

A leaked neural rendering library has spent the past week traveling backwards through four GeForce generations, and each stop has been slower than the one before it. Getting the runtime to load turned out to be a software problem; getting it to run at a usable speed is a hardware one. That gap is the useful part, because NVIDIA still has not said which cards DLSS 5 will officially support.

Ada needed a binary swap, not new silicon

What kept the leaked nvngx_dlssnr.dll on RTX 50 cards was not the model but the CUDA binaries compiled into the library for Blackwell. RTX Remix modder Uncle Burrito identified the incompatible ones, patched in Ada-compatible builds, and had Neural Rendering running on an RTX 4090 and an RTX 4080, as VideoCardz reported. Tom's Hardware verified the patch on an RTX 4080 and says other builds are in progress in the RenoDX Discord, where the runtime pulled out of NBA 2K27 has been circulating.

The cost lands roughly where Blackwell's does. Uncle Burrito's RTX 4090 test falls from about 135 FPS to 82 with Neural Rendering active, near 39 percent, against the almost 50 percent an RTX 5070 Ti gave up in Control. Ada has native FP8, which is what the model runs on, so nothing is being emulated here; the gap was only compilation targets that an early development build had no reason to include.

Ampere and Turing load it and little else

Below Ada the picture changes, because Ampere has no native FP8 and the FP16 path pays for the difference. An RTX 3070 Laptop running Kingdom Come: Deliverance II through an experimental SM86 async FP16 build went from about 60 FPS to a render latency of roughly 3,326 ms, up from 29 ms, with the standalone Neural Rendering test taking around 1.2 seconds per evaluation, per VideoCardz. Deep Rock Galactic on an RTX 3080 dropped from roughly 130 FPS to 4. RTX 3050 and RTX 3060 Ti testers reported around 1 FPS.

The better Ampere results are the informative ones, because of what they cost. Tom's Hardware's survey of the same Discord found GTA V at 20 to 30 FPS on an RTX 3080 Ti at 1440p and Dying Light 2 near 40 FPS on an RTX 3080 with a visibly noisy image, and the 41 FPS someone reached in Cyberpunk 2077 needed 720p on the ultra performance preset. Trading base resolution for neural fidelity is a strange bargain to strike on purpose.

Turing is a stage earlier again. ShortFuse has shared RTX 20 builds including one described as a speed boost, and VideoCardz reports that Red Dead Redemption runs while its sequel does not, Hogwarts Legacy sits in a failed state instead of activating, and one GTA V attempt lasted about ten seconds before the GPU stopped sending a display signal. Work there has moved on to profiling and optimization, which is a fair description of where those two architectures stand.

None of this constrains NVIDIA, which has committed only to a fall 2026 release and named no supported generations. Press coverage is further along: Linus Tech Tips has signed a DLSS 5 embargo agreement and has a video ready, confirmed on the WAN Show. DLSS 4.5 Ray Reconstruction officially covers RTX 20 through RTX 50, so a Blackwell-only launch would break with recent practice, but a patched DLL proving a card can execute a model is not the same as that card being supportable. FSR 4 on RX 7000 followed this shape exactly: an unofficial port first, then an official version AMD had to build properly.

The current driver blocks mVolt+ power limit control

The latest GeForce driver stops mVolt+ v0.36 from setting the core power limit on RTX 50 cards. The control blacks out as though the driver had crashed, whatever extra current is applied, according to a report from UNIKO's Hardware relayed by Tom's Hardware. Both give the version as 615.56, which matches nothing in NVIDIA's current sequence: Game Ready 616.56 shipped on August 26 and replaced 610.88, so the reported number reads as a transposition rather than a separate branch. What matters either way is that mVolt+ is the open-source tool that had been pushing an Asus ROG Astra RTX 5080 to 680W against a 450W OC BIOS ceiling, and RTX 5090s into the 650W to 700W range under load, with no VBIOS work or hardware modification.

Whether the block is deliberate is unconfirmed. Overclock.net users report Hydra 2.3B Pro still applying power limit changes on the same driver, which points at a compatibility break in the specific path mVolt+ uses, the NVVDD overcurrent limit, rather than a general lock on the interface. If that reading holds, the fix is an update from the tool's author rather than a driver rollback, and anyone unwilling to wait has a third option in Hydra.

Sources