MGPU Bridge v0.90 – FSR 3.1 Frame Generation

MGPU Bridge v0.90 — An experimental performance-focused fork based on Marcelo's Neural-coprocessor v0.2.5, featuring AMD FSR 3.1 Frame Generation, custom FG, LateLatch optimizations, improved GPU memory handling, and enhanced multi-GPU synchronization.

Highlights

AMD FSR 3.1.6 Frame Generation: Integrated using AMD's no-swapchain API with DComp presentation.

Custom Frame Generation: Retained as an alternative to AMD FSR for performance and image-quality comparisons.

LateLatch Optimizations: Improved neural-rendering frame handling and scheduling.

GPU Memory Improvements: Asynchronous staging and optimized cross-adapter transfers.

Improved Synchronization: Additional GPU fence tracking, buffer ownership protection, and safer shutdown handling.

Enhanced Performance HUD: Displays neural-rendering throughput, generated frames, and presentation FPS.

Multi-monitor DComp Support: Supports experimental dual-GPU configurations with multiple active displays.

NVAPI Reflex: Experimental low-latency integration.

Performance

In testing with an RTX 3080 Ti handling game rendering and an RTX 5060 Ti handling neural rendering and frame generation, the custom generator achieved approximately 124–130 presented FPS at 1440p output in tested scenes.

AMD FSR 3.1 may have different performance and image-quality characteristics.

Installation

Install the MGPU Bridge add-on according to the project README.

For AMD FSR Frame Generation, place amd_fidelityfx_framegeneration_dx12.dll in the game's mgpu folder alongside nvngx_dlssnr.dll.

Use the supplied mgpu.ini configuration for experimental frame-generation features.

Important

This is an experimental, unofficial fork. Compatibility and stability are not guaranteed. Live switching between frame-generation implementations is still undergoing validation.

Credit for the original MGPU Bridge project belongs to Marcelo Guibout. AMD FidelityFX components remain subject to their respective licenses.

MGPU Bridge 0.2.5 – LateLatch Optimized Fork

MGPU Bridge 0.2.5 – LateLatch Optimized Fork

Experimental performance-focused fork of Marcelo's Neural-coprocessor MGPU Bridge.

Improvements

• LateLatch optimization: Overlaps frame transfers and neural rendering to improve pipeline utilization and output throughput.
• Improved frame handling: Reduces unnecessary speculative transfers and stream synchronization overhead.
• Enhanced FPS HUD: Displays MGPU input, neural-rendered frames, presentation FPS, and performance diagnostics.
• Multi-display support: Allows explicit DComp overlay operation with multiple active displays.
• Improved synchronization: Additional fence handling, timer cleanup, and diagnostic improvements.
• Experimental optimizations: Includes further work on GPU transfer staging and Streamline diagnostics.
• **Optimized GPU memory transfers: Adds direct VRAM-to-VRAM transfer support to reduce unnecessary system-memory staging where supported, potentially improving transfer latency, memory efficiency, and frame pacing.

Performance observations

On an RTX 3080 Ti + RTX 5060 Ti configuration, testing showed neural-rendered throughput increasing from approximately 34 FPS to around 50–53 FPS with LateLatch enabled. Results will vary by system and game.

Notes

This is an experimental community fork, not an official release from the original developer. Features remain under development and may require configuration.

Original project: https://github.com/maohgad-web/Neural-coprocessor