RTX 4090 Undervolt Cuts Nearly 50 W From DLSS 5 With Almost No Performance Loss
NVIDIA's DLSS 5 Neural Rendering can place a surprisingly heavy power load on GeForce GPUs, but one RTX 4090 user has demonstrated that a carefully tuned undervolt can cut consumption by nearly 50 W while maintaining essentially identical gaming performance.
The test comes from Reddit user om_the_best_, who tested a GeForce RTX 4090 using the unofficial DLSS 5 implementation currently circulating through the enthusiast community. The user paired Neural Rendering with Multi Frame Generation and compared the GPU at its normal configuration against an established undervolt profile. The results show that DLSS 5 can significantly increase GPU power demand, but also that the RTX 4090 has considerable efficiency headroom when voltage is reduced.
In Hogwarts Legacy with DLSS 5 disabled and Multi Frame Generation running at 4X, the RTX 4090 consumed approximately 386 W while producing 360 FPS. Applying the undervolt reduced consumption to just 311 W while performance remained at approximately 355 FPS. That represents a 75 W power reduction for a difference of only 5 FPS.
The results become more interesting once DLSS 5 Neural Rendering is enabled. With DLSS 5 and 4X Multi Frame Generation active, the RTX 4090 consumed 435 W while delivering 175 FPS. After undervolting, consumption dropped to 388 W with the frame rate remaining exactly at 175 FPS. That is a 47 W reduction with no measured performance loss in the user's test.
Testing without Multi Frame Generation produced a similar result. DLSS 5 pushed the RTX 4090 to approximately 426 W while running at 51 FPS. With the undervolt applied, power consumption fell to 387 W while performance actually measured 52 FPS, although a difference of 1 FPS is well within normal benchmark variance.
The experiment is particularly relevant because Neural Rendering does not behave like a conventional graphics setting. DLSS 5 places additional AI workloads on the GPU alongside traditional rendering, and community testing has shown that a GPU reporting close to full utilization can consume substantially more power once Neural Rendering becomes active. One separate RTX test recorded GPU consumption increasing from 255 W to 336 W under otherwise heavily loaded conditions after DLSS 5 was enabled, alongside higher GPU and hotspot temperatures.
There is an important limitation to the RTX 4090 testing. DLSS 5 is not officially available for GeForce RTX 40 Series GPUs yet. NVIDIA has confirmed that official RTX 40 Series support is planned, but the current experiments use community modifications built around the available Neural Rendering runtime. Modders have already demonstrated DLSS 5 operating across RTX 4090 and RTX 4080 hardware, showing that Ada Lovelace can execute the workload even before NVIDIA releases its officially optimized implementation.
The same user is also running a community tool that enables DLSS Multi Frame Generation on RTX 40 Series hardware. The Universal RTX 40 MFG Unlock intercepts Streamline and NGX functions and exposes additional frame multipliers on compatible DirectX 12 and Vulkan games. The RTX 4090 user reported testing the implementation across titles including Crimson Desert, Hogwarts Legacy, 007 First Light, Cyberpunk 2077 and A Plague Tale: Requiem.
The efficiency improvement should not be interpreted as a universal RTX 4090 result. Undervolting depends heavily on individual silicon quality, frequency targets, cooling and workload behavior. A voltage setting stable on one GPU may crash another card, and different games can respond differently even when using the same profile. The result also does not mean undervolting eliminates the possibility of a power connector failure. It simply reduces the amount of power the GPU requires in this particular configuration.
What the experiment does show is that DLSS 5 may expose another reason for enthusiasts to revisit GPU efficiency tuning. Neural Rendering can substantially increase power consumption even when conventional GPU utilization already appears close to its maximum, making voltage optimization potentially more useful as AI processing becomes a larger component of real time graphics workloads.
The most interesting part of this test is not simply that undervolting saves power. RTX 4090 efficiency tuning has been popular since the GPU launched. What changes with DLSS 5 is the workload.
Neural Rendering puts additional pressure on parts of the GPU that traditional rasterization and even ray tracing do not exercise in exactly the same way. Seeing power jump from 386 W without DLSS 5 to 435 W with Neural Rendering active, while an undervolt pulls that figure back to 388 W, shows how much unused efficiency headroom can still exist in a flagship GPU.
For RTX 40 owners, the results are even more interesting because NVIDIA has not yet released its official Ada optimized DLSS 5 implementation. Community builds are already running, but the final NVIDIA version could behave differently in performance, power consumption and stability. This test should therefore be viewed as an enthusiast experiment rather than a universal optimization guide.
Would you undervolt your GPU specifically to reduce the additional power consumption from DLSS 5, or would you rather keep the card at its factory settings?
