Researchers build a PyTorch training network from Mac Studios

Researchers have introduced MCCL, an open-source backend designed to train PyTorch models across multiple Mac Studios connected by switched Ethernet. In evaluations across 24 nodes, they report high TCP goodput and parallel efficiency during Vision Transformer training, putting ordinary Mac infrastructure to work on distributed machine learning.

Watch Desk analysis

What happened

The paper, “MCCL: Training Pytorch models across a fleet of Macs at Skywalker Sound”, describes the Metal Communication Collective Library as an MPS backend for multi-node training. The researchers’ evaluation covers a 24-node Mac Studio setup and Vision Transformer training. They report high TCP goodput and parallel efficiency, but the supplied summary gives no specific benchmark figures.

Why it matters

Distributed training is often associated with specialist accelerator clusters. MCCL points to another possible route: connecting a fleet of Macs over Ethernet and using it for PyTorch training. That could give organisations with suitable Mac hardware a way to use more of their existing computing capacity, although the available evidence does not establish how its performance compares with dedicated GPU systems or other setups.

Our read

This is a concrete AI-computing development, not just a clever way to keep idle machines busy. The reported 24-node evaluation makes the idea more than a single-machine demonstration, while the absence of detailed figures means readers should treat “high” performance as the researchers’ characterisation, not a head-to-head result. The open-source approach is worth watching for developers who already have Macs and want to explore distributed training without starting from a conventional accelerator cluster.

What to watch

• Whether the project publishes detailed benchmark results and comparisons with other hardware.
• Which PyTorch workloads and Mac configurations MCCL supports in practice.
• Whether independent users can reproduce the reported training results.

Discussion spark: If a fleet of Macs can handle distributed PyTorch training, would you rather put existing machines to work or invest in specialist accelerator hardware?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.