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.
Chat and API
Jobs
Admins
StudentsBrowser, or scripts with an API key
AdminsPasskey sign-in
Cloudflare AccessOne-time email codes for chat.c4e.ai and the job portal
Lab tailnetTailscale, admins and machines only
Open WebUIChat, and a personal API key per student
Open OnDemandJob portal: files, job composer, active jobs
Admin toolsGrafana with Prometheus, Tailscale SSH
LiteLLMFixed model names, usage per student
Slurm schedules everything on the Sparks
Inference engineA Slurm job. TensorFold today; any OpenAI-compatible engine, such as NIM, swaps in.
Research jobsContainers through Pyxis and Enroot
Node 1DGX SparkGB10 · 128 GBOnline
Node 2DGX SparkGB10 · 128 GBOnline
Node 3DGX SparkGB10 · 128 GBComing soon
Node 4DGX SparkGB10 · 128 GBComing soon
Node 5DGX SparkGB10 · 128 GBComing soon
Node 6DGX SparkGB10 · 128 GBComing soon
CRS804 · rail AMikroTik 400G, split to 200G per Spark over ConnectX-7
CRS804 · rail BSecond ConnectX-7 port on every Spark
CRS312 · 10GbEMikroTik. Management, network boot, internet
OnlineComing soon
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
While there are only two Sparks, a research job pauses the served model, and Slurm brings the model back when the job ends.