Nvidia DLSS 5 mixed precision tests show limited gains on RTX 50 series

Mod developers testing Nvidia’s DLSS 5 neural rendering on Blackwell-based RTX 50 series hardware have reported limited performance gains when using mixed precision.

Experiments conducted via the OptiScaler project attempted to reduce inference costs by applying FP8 and NVFP4 mixed precision to DLSS 5 models. In tests using Baldur’s Gate 3 at 4K resolution, the transition to NVFP4—a low-precision data format native to Blackwell Tensor Cores—resulted in a frame rate increase of only approximately 1% to 2% during the neural rendering stage.

While the NVFP4 format is designed to reduce memory requirements and leverage hardware acceleration, the current improvements remain marginal. Developers noted that the mixed-precision solution does not cover all computational processes, as unsupported model shapes revert to FP8. Nvidia has previously stated that DLSS 5 performance has improved significantly since its initial demonstrations and that official support is planned for the RTX 40 series following its initial launch on RTX 50 series cards.