Two cards, one model
Splitting a model across two cards gives you more VRAM, not more speed — each token still crosses the link between them. It pays off when one card cannot hold the quant you want. Matching cards and a fast interconnect matter more than the total memory.