LLM across different generation GPUs (RDNA3 and RDNA4)?

A user on the Level1Techs forum asks whether local large language models can be run across both an AMD Radeon RX 7900 XT and a Radeon Pro R9700, noting the R9700's extra 12GB VRAM would bring combined capacity to 42GB, and questions if mixing GPU generations (RDNA3 and RDNA4) would work. The post highlights a common concern among AI enthusiasts about multi-GPU LLM inference with heterogeneous hardware.

Hondo https://forum.level1techs.com/u/Hondo 1 Simple question. Probably a complex answer. I have an 7900XT which I’ve dabbled with local LLMs on but I need more VRAM and will likely get a Radeon Pro R9700. The extra 12GB will be enough it’s never enough but I’d like to get a little boost by having the 42GB of both cards if it’s possible to run models and context across both. I’m new to this but I have a hunch trying to mix generations probably wouldn’t work well. Is that so?