Mapping GPUs to LLMs (and back): A bandwidth-based estimator for local inference
localllm-advisor.com