Higher Education AI Lab

A repeatable design for university AI teaching labs on Dell deskside hardware — multi-cohort teaching: coding agents, video search & summarization, and agent runtimes.

Summary

Universities need AI compute close to the classroom, manageable by campus IT, and strong enough for modern coursework. This template serves models in vLLM containers on Dell deskside workstations; students connect over the lab LAN with OpenAI-compatible clients, and course lifecycle is a fleet operation — one managed fleet serves many courses (N×M scheduling) with fast state reset between classes. It wraps three NVIDIA playbooks and cross-links the sibling solutions supplying the course modules: Agentic Coding (coding agents), Agentic Video Search & Summarization (video search & summarization), and OpenClaw on NemoClaw (agent runtimes; advanced modules via Hermes on NemoClaw).

Reference design. Pool sizing comes from measured ladders on the same hardware (Benchmarks & performance); your pilot observations tune it.

Personas

More: Personas.

Reference NVIDIA playbooks

The RTX Local AI hub covers the discrete RTX PRO options used by the M-tier shared amplifier.

Use cases

Teaching scenarios on a 30-seat reference lab:

Course profiles cover: notebook + local LLM · multi-model comparison (via LiteLLM) · fine-tuning (Unsloth Studio) · RAG (nomic-embed-text) · creative (ComfyUI + FLUX.1-dev / SDXL) · multimedia (Whisper large-v3).

What this template is not: hardware procurement; a curriculum (faculty pick content); a research-cluster substitute (L shared compute is classroom-scale).

Architecture

Usage and connectivity — students and devices reach the lab only through campus identity and one HTTPS entry point (view as published in the launch brochure):

Lab topology — workstations serve models over the lab LAN (doc-example range 203.0.113.0/24; real deployments use the campus address plan):

 Students: thin clients (S) · GB10 seats (M/L — the seat is the workstation)
        │  lab LAN 203.0.113.0/24 (doc example)
        ▼
 Management node — reverse proxy · LiteLLM router · JupyterHub
 (campus SSO/IdP · Landscape · Prometheus + Grafana :3000)
        │
 ┌──────┴──────────────────────────────┐
 ▼                                     ▼
 Model-serving workstations           Shared frontier pool (L only)
 GB10 pool / seats · optional         Dell Pro Precision with GB300 (~1:10)
 Precision 9 T4/T6 RTX PRO 6000             vLLM · public images
 (multi-instance or TP)
        │
        ▼
 NAS / NFS — weights · datasets · course profiles · home dirs
 (model proxy/cache · snapshots for between-class reset)

Shared infra (all tiers): NAS/NFS + model proxy/cache · Landscape fleet management · campus SSO/IdP (LDAP/AD) · optional remote access (proxy / jump host / VDI broker).

S / M / L sizing

Tiers differ by seat ratio and serving topology, not by stack (S/M/L Sizing).

All NVFP4. Rule of thumb (inherited from the measured Agentic Coding S-tier baseline): one GB10 comfortably serves 2–3 concurrent notebook students at interactive latency — a 30-seat lab wants 10–15 GB10s in the pool. App stack: JupyterHub, OpenWebUI, vLLM/Ollama on all tiers; Unsloth Studio, ComfyUI, OpenCode on M/L seats; LiteLLM routing on S/L pools.

Model recommendation

Per-course model choice, not one model for every course; pin the default and route alternatives through LiteLLM (one endpoint per student, routing by API key / course group).

• Default (S, all coursework): Nemotron 3.5 Lightning 30B A3B NVFP4 (card) — instruction, coding, RAG; runs on every tier. Reasoning model: clients need max_tokens >= 512. M/L shared serving default: Nemotron 3.5 Super VL; Lightning is the measured alternative.
• Alts, same API: S / coding-heavy CS — Qwen3.8 27B NVFP4 (card); M / shared amplifier — Qwen3.8-Flash-Next NVFP4 (card); L / frontier coursework — GLM-5.3-Flash NVFP4 (card), a 320B-total MoE on one 252 GB GB300.
• Comparison labs: Gemma 4 26B A4B and Gemma 4 31B (dense), side by side via LiteLLM. RAG: nomic-embed-text + the default chat model.
• VSS teaching module: Nemotron 3 Nano Omni (VLM) + Lightning (LLM), both local on public vllm/vllm-openai images — Agentic Video Search & Summarization.
• Creative / multimedia: FLUX.1-dev / SDXL (ComfyUI), Whisper large-v3.

Inference stack

vLLM (preferred) — OpenAI-compatible, public vllm/vllm-openai images (docs.vllm.ai). Ollama is the per-seat developer quick-start; LiteLLM fronts shared pools for multi-model routing. Reference: Inference Stack & Models.

Deployment recipes

Catalog: docs/recipes/ · Dell Deskside Agentic AI Solutions.

Commands in this guide run from the repository root (git clone this repo and cd into it).

recipes/gb10/vllm/docker-nemotron-3.5-lightning-nvfp4.sh --port 8000
curl -s http://localhost:8000/v1/models

JupyterHub / OpenWebUI / VSS classroom profiles run as inline Compose in the dev guide. Guides — pilot: Developer Deploy Guide · semester rollout: IT-Admin Guide.

Management environment

• Fleet: Canonical Landscape — golden image, patching, per-course profile push; rollouts run without SSH.
• Identity: campus SSO / IdP (AD, Entra ID, LDAP) with per-course group scoping; students never see a raw serving port.
• Monitoring: Prometheus + Grafana (:3000) + DCGM — GPU use, vLLM queue depth, storage growth, fleet health.
• Course lifecycle (N×M): container reset ~2 s (profile / model swap) · LVM snapshot restore ~30 s (home dirs) · autoinstall re-image 3–5 min (end of term).

Benchmarks & performance (validation)

Classroom sizing is derived from measured ladders on the same hardware rather than a replay of a 30-seat class: the S pool from the GB10 ladders (Lightning chat: ~44 tok/s per session alone, ~31 at four concurrent sessions), the L pool from the GB300 ladders (Super VL chat: ~74 tok/s per session at 32 concurrent). Reset times depend on your storage (a container reset takes seconds; see the IT-Admin Guide for snapshot restore and re-image). The measured Agentic Coding S-tier baseline (GB10, 200/200 chat, prior reference model; Agentic Coding) is kept below for pool sizing:

Cost & TCO

Indicative BOM cost ratios for a 30-seat lab (order-of-magnitude planning, not quotes):

Storage planning (30 seats, one term, indicative): weights 200–500 GB on NAS · datasets 100–200 GB/course · home dirs 20–40 GB × 30 · +25 % snapshot overhead → ~2–3 TB usable NAS. S is cheapest at the cost of contention; M removes contention at ~2× S compute; L adds fine-tuning and frontier coursework at ~1.5× M compute.

Dependencies & risk factors

• Dependencies: the three upstream playbooks; vLLM public images on GB10 / Precision / GB300; Hugging Face access (or a pre-seeded model proxy) for first weight pull; campus IdP; Landscape.
• Risk factors: model lineup refreshes with the open-weights cadence; undersized S pools degrade TTFT under concurrent use — plan from the inherited ladder and pilot observations.