Rule of thumb: parameters x bits / 8 for the weights, plus 20% for cache and runtime. Real needs depend on context length and the serving software. Rental uses Lambda on-demand rates for 80 GB and 180 GB GPUs.