I spent the last few days testing Llama3 on different GPUs, to find the cheapest cost per token. Spoiler: it's the Nvidia L4, surprisingly.