Strix Halo offers the option to split memory to create a contiguous framebuffer for the iGPU for better performance (less data fragmentation), but iGPU can access the entire UMA region dynamically. So, in a 128GB system where 96GB is allocated, iGPU isn't locked out of the remaining 32GB - it's just dynamically managed between CPU and iGPU when large datasets are processed in RAM.
Once iGPU takes too much RAM, CPU processes spill into a pagefile, so there isn't a magic bullet for everything. You can allow the iGPU to use as much as needed via driver memory management, but that has other costs.