With Ollama i got the 20B model running on 8 TitanX cards (2015). Ollama distributed the model so that the 15GB of vram required was split evenly accross the 8 cards. The tok/s were faster than reading speed.