Skip to content
MacLustr
ENFR

Independent research infrastructure

Research at the edge of Apple Silicon.

A personal Apple Silicon research cluster, built to explore what distributed computing can make possible.

MacLustr is an evolving distributed computing infrastructure supporting experiments in local language models, machine learning, and computational research.

Explore the infrastructure Research collaboration
Black steel shelving unit in a room, with nine Mac Studio computers on three shelves and three Mac mini computers on the lowest shelf, network and power cables tied along the frame, two small monitors and a fan.
Shelving unit A, photographed on 7 November 2026

Measured, not estimated

Inventory verified on

  • 29 machines documented

    28 answered the latest discovery

  • 1,068GB of unified memory

    sum over separate machines, not one shared pool

  • 393 CPU cores

    GPU core total withheld: one record is incomplete

  • 1 GbE wired link per machine

    negotiated rate read on each node

Totals are computed from the node records below. Core and memory figures are hardware specifications read from each machine; they are not performance results.

Built from real machines

Three steel shelving units, one machine at a time.

I built MacLustr gradually, starting from a few Mac Studio computers and adding Mac mini and laptop nodes as research needs grew. It lives on ordinary industrial shelving, not in a data centre.

Nothing on this site is rendered or staged. The photographs show the cluster as it runs: labelled machines, tied cable runs, floor fans pushing air up through the shelves, and laptops working with their lids open.

See all photographs

Front view of a black shelving unit: two Mac Studio computers on top, a row of four Mac mini computers, a row with one Mac Studio and three Mac mini computers, and three black fans at the bottom; cables hang behind the shelves.
Shelving unit B — Mac mini rows and cooling

Topology

How the machines are connected

A simplified map. It groups machines by kind and shows roles, not addresses; the internal layout is deliberately left out.

Select a block to read what it does.

Boundary of the cluster
  • Local wired network
  • Encrypted tunnel

04 / 08

Wired Ethernet

Every machine reports a 1 GbE negotiated link. This is the measured rate; no faster interconnect is claimed.

Read the technical architecture

Hardware

Four kinds of machine

Each family leads to its individual node records.

Browse the inventory

Research themes

What the cluster is used for

  • 01Active

    Local language models

    Running and studying open-weight models on machines I control: architecture experiments on small models, and inference on larger ones within a single node's memory.

  • 02Active

    Apple Silicon and MLX

    Measuring what unified memory and the Apple GPU can do for numerical work, with MLX for models and Metal for custom kernels.

  • 03Active

    Distributed workloads

    Statistical methods designed for many machines and few communication rounds, tested on real hosts rather than simulated ones.

  • 04Active

    Scientific computing

    Large batch computations in econometrics and data analysis, split into independent jobs and placed in proportion to each node's cores.

  • 05Active

    Agentic research

    Autonomous coding agents operate the cluster through a single gateway, and are themselves a subject of study.

  • 06Planned

    Reproducible benchmarking

    A fixed protocol for comparing chip generations on identical models and prompts. The protocol is published; the measurements are not yet.

Selected work

Experiments that ran here

Each entry states how the cluster was actually used, including when it was a single machine.

  1. Working Paper 40 ·

    RAID: A Metal-Accelerated C++ Framework for Reproducible High-Frequency Financial Econometrics

    Use of the cluster — One experiment of the paper runs the same workload across eleven machines of the cluster to compare chip generations.

  2. Working Paper 41 ·

    Multi-Resolution Context in Intraday Probabilistic Return Forecasting: A Pre-Registered, Matched-Budget Measurement on 300 U.S. Instruments

    Use of the cluster — Training and evaluation ran on a single Mac Studio M3 Ultra node: one machine, not a distributed run.

  3. Working Paper 28 ·

    Beyond One-Step Distributed Maximum Likelihood: Local Bias Cancellation, Global Synchronization, and the Price of Inexact Curvature

    Use of the cluster — This is the cluster's clearest case of real distributed parallelism: the Monte Carlo study ran on 18 nodes and the real-host testbed used 16 machines exchanging messages over TCP.

Open the research log Benchmark methodology

Behind the scenes

Photographic journal

  • Black steel shelving unit in a room, with nine Mac Studio computers on three shelves and three Mac mini computers on the lowest shelf, network and power cables tied along the frame, two small monitors and a fan.
  • Second view of the black shelving unit with Mac Studio and Mac mini computers, labels with QR codes in front of each machine, bundled cables and a reel of yellow network cable.
  • Front view of a black shelving unit: two Mac Studio computers on top, a row of four Mac mini computers, a row with one Mac Studio and three Mac mini computers, and three black fans at the bottom; cables hang behind the shelves.
  • Black shelving unit with two MacBook Pro laptops on the top shelf, one showing a node status screen, two MacBook Air laptops on the second shelf beside storage enclosures, three Mac mini computers on the third shelf and fans at the bottom.

Open the gallery

Collaboration

Interested in using MacLustr for research?

MacLustr primarily supports my own research. I am open to discussing collaborations and, where capacity and project fit permit, access for external researchers. Requests are reviewed individually; access is not guaranteed.

How to propose a project contact@spboucher.ai