Independent testing of a leaked version of the DLSS 5 neural filter has revealed an unexpected hardware limitation of the flagship GeForce RTX 5090 graphics card. version Tom’s Hardcore We tested the technology in games like Cyberpunk 2077, Hogwarts Legacy, and Control, comparing a reference RTX 5090 Founders Edition to an overclocked MSI RTX 5090 Lightning Z. The results show that the bottleneck of heavy neural network models is not the computing power of the graphics card, but the limitations of the power subsystem.
The test uses OptiScaler injection based on the leaked DLSS 5 library, and the overhead of processing a frame is exactly consistent with NVIDIA’s own statement of about 8 milliseconds per frame.
RTX 5090 Founders Edition (single 12V-2×6, 575W limit) – Enabling DLSS 5 resulted in an average frame rate drop of 42-49%. At the same time, the graphics card instantly reached the hard power consumption limit of 575W and reset the chip’s operating frequency.
MSI RTX 5090 Lightning Z (two power connectors, maximum limit of 1000 W) – Custom board with two cables showed significantly fewer framerate drops, and the metric for rare events was significantly more stable (1% lower). However, energy consumption is increasing at an alarming rate. In Hogwarts Legacy, for example, the accelerator’s power jumps to 720 W simply due to the activation of the neural filter.
Obviously, the standard version of the RTX 5090 with a single 16-pin power connector is hardware limited in terms of TDP and cannot achieve full operation of a neural network of this scale. Only advanced graphics cards with two connectors or greatly increased power limits can unlock the full potential of DLSS 5.
