If you can't, your performance will likely be abysmal though, so there's almost no middle ground for the LLM workload.