---
title: "Architecture"
description: "How the C4E cluster fits together, from sign-in to the six DGX Sparks and their 200 Gb/s fabric."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.c4e.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Architecture

Six NVIDIA DGX Sparks on a shared 200 Gb/s fabric. Together they have 768 GB of unified memory and about 1.6 TB/s of aggregate memory bandwidth (6 × 273 GB/s). Students reach chat and the job portal through an email-code sign-in; admins come in over a private tailnet.

## Sign-in

Students sign in with a one-time code sent to their email. There are no passwords to keep, and nobody installs a VPN.

## Models

Each model runs as a Slurm job on whichever engine is fastest for it. LiteLLM keeps the model names fixed, so an engine swap changes nothing for students or their keys.

## Sharing the Sparks

The model runs as a Slurm job too, in an admin-only `serve` partition at the lowest priority. Student and research jobs go to the default `jobs` partition, which ranks higher. When a job needs the model's nodes, Slurm preempts the model: it stops it and puts it back in the queue (preemption mode `REQUEUE`). The job starts at once.

While there are only two Sparks, the model spans both, so almost any job pauses it. From the moment the job starts until the model has reloaded, chat.c4e.ai and api.c4e.ai can't answer and requests fail with an error. When the job ends, the model starts again by itself and needs a couple of minutes to load. A 14-second test job kept it offline for about two minutes.

Students can see this with `squeue` and `sacct`, and can keep the pause short by asking only for the `--time` they need or by scheduling long jobs with `--begin`. [Slurm jobs](https://docs.c4e.ai/getting-started/slurm-jobs/#jobs-pause-the-shared-model) has the details. Once the four new Sparks arrive, the model gets nodes of its own and jobs stop pausing it.

Source: https://docs.c4e.ai/architecture/index.mdx
