SiliconBench benchmarks nine LLM serving engines on Apple Silicon unified-memory desktops across speed, memory headroom, and output fidelity rather than speed alone. Testing Qwen3, Qwen3.5, and Gemma 4 under concurrent chat and agent workloads, the authors find that explicit memory budgets do not guarantee headroom, that only three of nine stacks satisfy completion, fidelity, and model-coverage gates together, and that one engine’s packed prefill-decode path maintains lower first-token latency under load. They also report that tensor parallelism over Thunderbolt RDMA scales across two machines while pipeline parallelism over TCP regresses.
