Unusual parallel inference using consumer RTX rig

I am a local LLM novice and I dont have that much knowledge about local inference, but I have been a gamer for years and I know nVidia gpus since Riva TNT, I also know my rig quite well, my idea is to utilise redundant otherwise iGPU

This report outlines the design and implementation of The Sentinel Module —a dedicated, out-of-band monitoring system designed to act as a high-reliability guardian for your primary LLM pipeline. By isolating this module onto the integrated GPU (iGPU) with a specific 8GB memory allocation, we create a “fail-safe” layer that ensures the integrity of the Hermes agent without consuming the resources or performance overhead of the main inference engine.