Skip to content
MacLustr
ENFR

Research log

What has been done, what is under way, what is proposed

Entries are listed with their status. Completed work links to its manuscript; ongoing and proposed work makes no result claim. For each entry I state how the cluster was used.

Completed

  1. Working paper

    • Scientific computing
    • Apple Silicon and MLX
    • Reproducible benchmarking

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

    A C++20 framework that runs high-frequency econometric estimators on Apple GPUs through Metal, with double-float arithmetic to work around the absence of 64-bit floating point on the GPU, and measures where the GPU actually beats the CPU.

    Use of the cluster

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

    Reference

    Simon-Pierre Boucher (2026). RAID: A Metal-Accelerated C++ Framework for Reproducible High-Frequency Financial Econometrics. Research Working Papers, Working Paper 40. Read the paper

  2. Working paper

    • Machine learning
    • Scientific computing

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

    A pre-registered comparison of a multi-resolution patch transformer against matched-budget baselines for 5- to 60-minute probabilistic return forecasts.

    Use of the cluster

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

    Reference

    Simon-Pierre Boucher (2026). Multi-Resolution Context in Intraday Probabilistic Return Forecasting: A Pre-Registered, Matched-Budget Measurement on 300 U.S. Instruments. Research Working Papers, Working Paper 41. Read the paper

  3. Working paper

    • Distributed workloads
    • Scientific computing

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

    Statistical theory for estimators computed on many machines and combined in a few communication rounds, with Monte Carlo evidence and wall-clock measurements on real hosts.

    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.

    Reference

    Simon-Pierre Boucher (2026). Beyond One-Step Distributed Maximum Likelihood: Local Bias Cancellation, Global Synchronization, and the Price of Inexact Curvature. Research Working Papers, Working Paper 28. Read the paper

  4. Working paper

    • Local language models
    • Machine learning

    PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models

    An architecture study for small language models with a sparse workspace trained under counterfactual supervision. The main result is negative and reported as such: learned routing did not beat the no-routing baseline.

    Use of the cluster

    Training runs were scheduled across cluster nodes as independent jobs, alongside a laptop.

    Reference

    Simon-Pierre Boucher (2026). PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models. Research Working Papers, Working Paper 31. Read the paper

Ongoing

  1. Experiment

    • Local language models
    • Apple Silicon and MLX

    Expert orchestration for mixture-of-experts models on MLX

    A runtime study of how mixture-of-experts language models can be laid out and scheduled on Apple Silicon. The manuscript is not public yet, so no result is quoted here.

    Use of the cluster

    Runs on individual nodes with MLX. Cross-node experiments are constrained by the 1 GbE links measured on every machine.

    Reference

    No public manuscript yet.

Proposed

  1. Protocol

    • Reproducible benchmarking
    • Local language models
    • Apple Silicon and MLX

    A reproducible local-inference benchmark across Apple Silicon generations

    Planned: the same open-weight models, quantizations and prompts measured on every chip family in the inventory, from M1 Max to M6, following the protocol published on the benchmarks page. No measurement has been published yet.

    Use of the cluster

    Would use one node per chip family, one at a time, with nothing else running on the machine.

    Reference

    No public manuscript yet.

Software stack

What the experiments above are built with.

Models
MLX and mlx-lm for language models on Apple GPUs; PyTorch for conventional training
Numerics
C++20 with Metal for custom GPU kernels; Python scientific stack for analysis
Data
Columnar files on local SSD, PostgreSQL on the node that needs it
Operations
A custom orchestrator, a process manager per node, git as the source of truth

Ownership and affiliation

MacLustr is personally owned, funded and operated. It does not belong to, and is not funded or endorsed by, any university or other institution. Papers listed here carry their own author and series information.