AMD GPUs Can Now Run DLSS 5 Through a New Mod

DLSS 5 is changing how neural rendering works across GPUs, while new mods are extending the technology beyond its intended hardware. At the same time, dual-GPU rendering is being tested to handle the performance cost of neural rendering, and new memory research could ease some of the supply pressure created by AI demand.

AMD GPUs can now run DLSS 5. NVIDIA's new technology has already been tested on older NVIDIA GPUs from the 40 series down to the 20 series. It has also been tested in games that do not officially support it, and all of this happened before AMD released the new technology. Now that DLSS 5 is officially here, a new mod allows RX 9000 and 7000 GPUs to run DLSS 5.

AMD GPUs Can Now Run DLSS 5

Many people think of NVIDIA as a hardware company, and it is, but hardware isn't what gives NVIDIA a near monopoly in the AI space. It is the software, specifically the CUDA suite of software. This is similar to what makes Apple products compelling. Think iMessage, Apple Pay, or iCloud. It is called a walled garden, and once you're in it, it can be hard to leave.

This is what NVIDIA hopes to do with DLSS 5. To some extent, previous DLSS releases worked in the same way, but DLSS 5 changes the situation. Instead of giving the user more performance or slightly better lighting, DLSS 5 changes the scene altogether. Some people don't like the look, but DLSS 5 could encourage developers to use it as a design crutch.

Developers could one day expect the user to have DLSS 5 turned on because their character design or other design is poor without it. Then users would basically have to use NVIDIA GPUs, which is the company's hope. That is why NVIDIA is unlikely to be happy about this mod.

DLSS NR Brings DLSS 5 to AMD GPUs

The mod is called DLSS NR on AMD. Ancient Gameplays has already tested it, and it works. The mod only requires a few basic steps to get running, but it has one issue: performance. After testing a 9070 XT in Cyberpunk 2077, with DLSS 5 off, it averaged around 85 FPS. As soon as DLSS 5 was turned on, the FPS dropped to 11 to 12 FPS.

The mod developer offers some hope. According to the developer, the 9070 XT has reached 33 FPS at 1080p, and performance is being improved to match a 5070 Ti. The developer believes AMD cards can run DLSS 5 with a performance loss similar to NVIDIA's 50 series cards. Time will tell if NVIDIA tries to stop this, but the development could give AMD GPU owners another way to use DLSS 5.

Game support was also inconsistent, with micro stuttering and sometimes barely any extra performance. The new dual-GPU setup works differently. It offsets the performance hit from DLSS 5. DLSS 5 can have a large effect on FPS, but a new mod called Neural Coprocessor allows one GPU to render the game while the second handles DLSS 5's neural rendering.

Dual-GPU Setup Still Has Limitations

This clearly makes a difference, but the current setup has problems. The FPS numbers measure how fast the game is running on the first GPU. That doesn't necessarily show how fast the second GPU displays the finished AI-enhanced image. If the second card cannot keep up, the mod can skip older frames and move on to newer ones.

The game might report 91 FPS, but that does not guarantee that 91 FPS with neural rendering is actually being displayed. To determine how much of the improvement reaches the screen, you need to measure the final output. The recommended setup also requires two monitors, with one connected to each GPU.

Running the second GPU without its own monitor reportedly reduced throughput by roughly 33%. It also adds latency, although the full input-to-screen delay has not been measured. The technology is still experimental. If these problems can be addressed and DLSS 5 becomes popular, dual GPUs could return.

The goal is to pack more memory into existing fabrication facilities.

Kepler Computing says its memory is designed to offer capacities beyond conventional HBM and SRAM while approaching SRAM's bandwidth-per-watt. The goal is more capacity with efficient data movement, which is the type of memory technology AI companies need.

The US government has also announced proposed support of up to $245 million for Kepler's memory research. That funding is separate from the $468 million in private funding. The timing remains important. Kepler says it plans to produce samples before the end of this year, with production ramping in 2027.

Producing samples and shipping enough memory to make a difference in global demand are two different things. The company still has to prove that it can manufacture the technology at scale. If the technology works out, it could mark the beginning of the end for the current memory crisis.