A cluster of separate computers

01

Unified memory, one machine at a time

On Apple Silicon the CPU and GPU share one pool of memory. A model that fits in a node's unified memory can be loaded once and used by the GPU without copying it to a separate graphics card. This is why a 96 GB Mac Studio can hold models that would not fit on most consumer GPUs.

That pool stops at the edge of the machine. Connecting computers with Ethernet does not combine their unified memory into a single shared GPU address space, whatever the link speed. The memory total on this site is a sum over separate machines: the largest model the cluster can load is bounded by its largest single node, not by the total.

02

Three kinds of workload

Independent: most work runs on one node from start to finish, such as training a small model, estimating a statistical model, or serving an application. The cluster then acts as many computers that can be used at once.

Scheduled across nodes: a large batch is cut into independent jobs and distributed in proportion to each node's core count. Nodes do not talk to each other during the run; results are collected at the end. Most multi-node computation here is of this kind.

Truly distributed: a few experiments have nodes exchange messages during the computation, for example distributed maximum-likelihood estimation with a small number of communication rounds over TCP. These are specific research protocols, not a general capability.

No model has been split across machines here. I make no claim of model sharding or of distributed inference, and training in the usual sense has been limited to models that fit on one node.

03

Network

Every machine that reported a wired link negotiates 1 GbE (1000BASE-T). I state the measured rate rather than the rated capacity of any switch or adapter. A faster interconnect is not in service on the nodes, and none is claimed.

At 1 GbE the network is the bottleneck for anything that moves tensors between machines: a single generated token is cheap to send, but activations or model weights are not. This is the main reason the cluster favours independent and scheduled jobs.

04

Storage

Working data lives on each node's internal SSD, between 245 GB and 995 GB per machine as reported by the machines. Large datasets are kept on the node that computes on them, to avoid reading them over the network.

Two network-attached storage units provide backups and shared files. They are not used as a parallel filesystem.

05

Orchestration and control plane

One Mac Studio is the gateway. It hosts a private git service and an orchestrator that I wrote: it scans each node's free memory, load, disk and installed runtimes, scores the candidates, places a workload on the best one, checks its health and records where it runs.

A repair loop runs every few minutes and restarts what is missing after an interruption. Machines are configured to log in and resume on their own after a power cut; this was tested involuntarily in September 2026, when sixteen machines restarted together and came back unattended.

This describes how the system is built. It is not an uptime commitment: the cluster is a research installation and can be stopped, rewired or reassigned at any time.

06

Model serving

Open-weight models are served with MLX on a single node, started for the duration of an experiment. An always-on, OpenAI-compatible endpoint that served several models existed earlier and was dismantled in June 2026; it is listed in the history, not as a current service.

No public inference endpoint is offered, and none of the node addresses is reachable from the internet.

07

External boundary

One hosted bare-metal server sits outside the cluster. It terminates TLS for the public websites and relays requests to the right node through an encrypted tunnel. It does no research computation and is not counted in the inventory.

The site you are reading is served this way: by one node of the cluster, behind that gateway.

08

Power and cooling

The machines sit on open steel shelving at room temperature, with floor fans moving air up through the shelves, as the photographs show. There is no dedicated cooling plant.

Power draw, temperatures and energy per task have not been measured systematically yet. Until they are, this site publishes no power or efficiency figure.