AMD GPUs Split Local LLM Workloads

The real answer is “yes, sort of” for capacity, and “no” for clean pooled-memory performance. If you already own both cards, the smarter setup is usually workload separation on the 6800 XT and the big model on the 7900 XTX, not expecting a free upgrade in speed.

• –llama.cpp supports multi-GPU split modes and tensor splitting, so mixed cards can cooperate on one model
• –That cooperation has overhead; if a model fits on one GPU, splitting it across two usually hurts throughput
• –The most useful pattern here is division of labor: embeddings, reranking, memory management, or a draft model on the smaller card
• –Power is the other constraint: 850 W may work with aggressive power limits and undervolting, but there is not much headroom
• –This is an infrastructure story, not a model story; the win is fitting larger local models, not magically doubling performance