Note: Predicted prefill/decode throughput values (tokens/sec) are purely theoretical and subject to real-world constraints. Older cards may have worse software support. Some newer cards lack tensor cores data, making prefill estimates unreliable.