DLSS 5 Modding Community Develops Dual-GPU Workaround and Sub-Native Rendering Optimization
Following DLSS 5's official debut in NBA 2K27, PC modders have created two performance-focused modifications: Neural Copressor shifts neural rendering to a second GPU for dramatic frame-rate gains, while Neural Upstream reduces computation cost through earlier pipeline processing.

DLSS 5 has entered the modding community's hands following its official release through NBA 2K27. The technology's .DLL file spread rapidly across other PC titles, prompting developers to explore optimization approaches that address the feature's computational demands.
The first notable mod, Neural Copressor, developed by Marcelo Guibout, redistributes neural rendering workload across dual graphics cards. In testing with The Blood of Dawnwalker on a dual RTX 5060 Ti setup, the mod achieved up to 127 percent faster performance. However, this approach introduces a significant tradeoff: display latency doubles. The rendered frame must travel from the first GPU through the second GPU's neural processing pass before reaching the display—a dynamic similar to frame generation technology. Guibout suggests this architecture suits turn-based or timing-insensitive titles better than action-focused games.
A second optimization, Neural Upstream by Matias Lombo, takes a different approach by sacrificing some visual fidelity to improve frame-rates. Rather than performing neural processing on the full-resolution image near the pipeline's end, Neural Upstream processes at an earlier stage before DLSS upscaling occurs. This reduces computational overhead and enables higher frame-rate delivery while maintaining playability-focused performance gains.
Both mods represent the community's attempt to balance DLSS 5's visual capabilities against performance constraints on consumer hardware, offering different solutions depending on use case priorities.
Sources
- Digital FoundryEstablished publication · recorded Sep 8, 2026DLSS 5 Mod Shifts Load to Second GPU, Doubling Frame-Rates and Latency
