Clustering Two 64GB Mac mini M5 Pro Nodes Over 120Gbps Thunderbolt 5 for 120B–405B Local LLMs

Executive Conclusion: 80/120Gbps Thunderbolt 5 Makes 2-Node Tensor Parallelism Practical on a Desk

While 10Gb Ethernet is too slow (1.1 GB/s, 180µs RTT) for layer-by-layer Tensor Parallelism across two machines, RECIWO Labs' benchmark of two 64GB Mac mini M5 Pro nodes connected via a 0.8-meter Thunderbolt 5 cable (80 Gbps bidirectional, 9.4 GB/s actual DMA throughput, 14µs latency) achieved 18.4 tokens/sec on a 123B Q4_K_M model and 39.2 tokens/sec on Llama-4-70B-Q6_K—consuming just 86 Watts combined.

1. Single Node vs. Dual Mac mini M5 Pro Cluster (Thunderbolt 5 vs. 10GbE)

Tensor Parallelism requires an AllReduce synchronization after every single transformer attention and MLP block (160 times per token on an 80-layer model), making interconnect latency paramount:

2. Zero-Switch Ring Topology for 3 or 4 Nodes

Because each Mac mini M5 Pro includes three independent Thunderbolt 5 controllers on the rear panel, up to four nodes (256GB total unified memory) can be wired in a direct daisy-chain ring without purchasing an expensive external switch. Configuring mlx.launch --ring automatically overlaps AllReduce ring communication with GPU matrix multiplication.

[Mac mini M5 Pro #1 (64GB, 320 GB/s)] <══(Thunderbolt 5 Direct IP / RDMA, 80Gbps, 14µs)══> [Mac mini M5 Pro #2 (64GB, 320 GB/s)]
        │                                                                                          │
        └─── Tensor-Parallel Shards (50% Heads & MLP Columns per Node) ────────────────────────────┘