Which GPU should you actually run?

Answer five questions about your workload. The recommendation runs the same logic we use when scoping dedicated environments — memory footprint first, then interconnect, then cost per useful token — and sizes the cluster, not just the card.