LLM-Leistung im Jahr 2026: Benchmarks, Engpässe und Optimierung A performance engineering hub for running LLMs efficiently: runtime behavior, bottlenecks, benchmarks, and the real constraints that shape throughput and latency.