Gaiel Series 3-Mac Cluster Benchmark Report

CLUSTER_3NODE_BENCHMARK_REPORT.md

🚀 Gaiel Series 3-Mac Cluster Distributed Inference Benchmark Report

Hardware Cluster:

• 💻 MacBook Pro (Apple M5 Pro, 24 GB Unified Memory)
• 💻 MacBook Air (Apple M5, 24 GB Unified Memory)
• 🖥️ Mac mini (Apple M4, 16 GB Unified Memory)
• Total Unified RAM: 64 GB
• Interconnect: Thunderbolt 4/5 Low-Latency Bridge + High-Speed LAN

---

1. 3-Node Distributed Cluster Performance Summary

Model	Sharding Strategy	Prompt Processing (Prefill)	Generation Speed (TPS)	Peak Cluster Memory	Cluster Power
🥇 Gaiel-1.5B-Korean-Tuned-MLX	3-Node MlxRing Pipeline	307.57 tokens/s	22.11 tokens/s	0.78 GB	81.9 W
🥈 Gaiel-7B-Korean-Tuned-MLX	3-Node MlxRing Pipeline	192.79 tokens/s	16.27 tokens/s	2.12 GB	81.3 W
🥉 Gaiel-32B-Korean-Tuned-MLX	3-Node MlxRing Pipeline	46.95 tokens/s	4.10 tokens/s	7.96 GB	93.7 W
⚡ Gaiel-32B (Thunderbolt 2-Node)	2-Node MlxRing Pipeline	54.71 tokens/s	3.40 tokens/s	8.79 GB	69.4 W

---

2. 3-Node Hardware Telemetry Under Load (Gaiel-32B)

Node	Hardware Spec	GPU Utilization	Temperature	Power Draw
MacBook Pro	Apple M5 Pro (24GB)	79% (0 ~ 100%)	62.4 °C	52.5 W
MacBook Air	Apple M5 (24GB)	26% (8 ~ 36%)	44.4 °C	16.3 W
Mac mini	Apple M4 (16GB)	34% (16 ~ 62%)	76.8 °C	26.3 W

---

3. 72B Model (39.0 GB) 메모리 분석 및 권장사항

• 현재 3대 총 가용 RAM: 37.8 GB (시스템 기본 점유 제외)
• 72B 모델 로딩에 필요한 최소 RAM: 39.0 GB (약 1.2 GB 부족하여 Placement 대기)
• 해결 방법: 맥북 프로나 맥북 에어에서 무거운 웹 브라우저 탭이나 다른 백그라운드 앱을 정리하여 1~2GB 메모리를 확보하면 Exo가 72B 모델을 3대에 즉시 분산 배치하여 구동할 수 있습니다.