The National Research Platform
for AI-Enabled Research

GPUs and AI for research and education

Daniel Diaz · San Diego Supercomputer Center, UC San Diego

UCLA Health IT Research Tech Talk · 10 November 2026

Agenda

What is NRP?
Getting access, and what a namespace is
GPUs: what's available, and how to run a GPU job
AI services: hosted LLMs and Jupyter AI
JupyterHub for labs and classes
Live demo
Questions

Slides, the demo notebooks and the GPU job:

https://training.nrp-nautilus.io/ucla-health

The Platform

What is NRP?

A national Kubernetes cluster, Nautilus, for research and education
Run by SDSC (UC San Diego), Nebraska–Lincoln and MGHPCC; funded by NSF
Hardware contributed by 70+ institutions, UCLA among them
Scale today:
536 nodes
1,411 GPUs in 33 models
Petabytes of storage
Free for non-commercial research and education

https://dash.nrp-nautilus.io

Site locations from the NRP dashboard, 10 Nov 2026
Not drawn: 8 sites in Hawaii, Guam, Korea, Brazil and Europe

Kubernetes and Containers

Container image: your code plus everything it needs (Python, PyTorch, CUDA), in one package
Runs the same on a laptop, a cluster or the cloud
Kubernetes runs containers across many machines
You say what you need (CPUs, memory, a GPU); it finds a machine that has it
NRP is one Kubernetes cluster: 536 nodes

Linux

CUDA libraries

PyTorch, NumPy

train.pyyour code

Container image

built once, from a Dockerfile

Kubernetes cluster

Node64 CPU cores

Node8 × A10 GPUs

Node4 × A100 GPUs

the scheduler picks a node with a free GPU

your container

Access Methods

You

Hosted LLMsbrowser chat or an API key

AI gatewaychecks your token, routes

Model serversopen-weight models on NRP GPUs

JupyterHublog in, pick an image and a GPU

Your notebook servera container started for you

A CPU or GPU nodechosen by Kubernetes

Kuberneteskubectl apply -f job.yaml

Schedulerfinds a node with what you asked for

Your podsrun, write results, exit

No login node and no batch queue: pick the door that matches how much infrastructure you want to touch.

Most people start with JupyterHub or the hosted LLMs.

from openai import OpenAIclient = OpenAI(base_url="https://ellm.nrp-nautilus.io/v1",                api_key=TOKEN)        # from nrp.ai/llmtokenclient.chat.completions.create(model="gpt-oss",    messages=[{"role": "user", "content": "Hi"}])

In the browser at nrp-openwebui.nrp-nautilus.io, or from any code that speaks the OpenAI API

Log in with your campus account through CILogon
Choose an image, CPUs, memory and a GPU type
NRP's hubs: jupyterhub-west.nrp-nautilus.io and jupyterhub-east.nrp-nautilus.io
Today's demo: jh-training.nrp-nautilus.io

Your jobs

$ kubectl apply -f sweep/job.batch/train-lr-0.1 createdjob.batch/train-lr-0.01 createdjob.batch/train-lr-0.001 created$ kubectl get podsNAME                   READY   STATUS    RESTARTS   AGEtrain-lr-0.1-7xkqp     1/1     Running   0          40strain-lr-0.01-m4q9d    1/1     Running   0          40strain-lr-0.001-c8w3n   0/1     Pending   0          40s$ kubectl get podsNAME                   READY   STATUS      RESTARTS   AGEtrain-lr-0.1-7xkqp     0/1     Completed   0          3m10strain-lr-0.01-m4q9d    1/1     Running     0          3m10strain-lr-0.001-c8w3n   1/1     Running     0          3m10s

GPU node · 4 × A10

lr-0.1lr-0.001

lr-0.01

busy

GPU node · 4 × RTX 3090

busy

lr-0.001 waits in Pending until a GPU frees up; a finished job releases its GPU

What NRP Is For

Public or synthetic data

public datasets, simulations, benchmark and synthetic data

De-identified data

if your IRB and data use agreement allow it

Protected health information

PHI or identifiable patient data, including in LLM prompts

OK on NRP

Not on NRP

Good uses

model development and training · image analysis · simulation · teaching · LLMs on non-sensitive text

Also not for: commercial work, or services that need guaranteed uptime (it is a shared, best-effort cluster)

NRP is not HIPAA compliant. Keep PHI in UCLA Health's approved environments.

Getting Access

1. Log in at nrp.ai

2. Get into a namespace

Request one at get-started.nrp-nautilus.io
Or ask an existing namespace admin to add you

3. Start working

JupyterHub, kubectl, or an LLM token from nrp.ai/llmtoken

Who can request a namespace: faculty, staff, postdocs and research scientists. Students: your PI or instructor requests it and adds you.

https://get-started.nrp-nautilus.io

What is a Namespace?

Your group's workspace on the cluster
One per lab, project or course; the PI or instructor is its admin and adds members
Needed for GPU jobs, storage, your own JupyterHub, and API access to the hosted LLMs
Not needed to chat with the LLMs in the browser
What's inside is visible only to its members

Nautilus cluster

lab-smith

Dr. Smith (admin) and 3 studentsGPU jobs · a data volume · LLM tokens

bio-course

the instructor (admin) and 40 studentsa course JupyterHub · shared datasets

another-lab

their members onlyyou can't see their jobs or data

Talking to the Cluster

Your laptopkubectl + ~/.kube/config

CILogonyour UCLA login, once

Nautilus APIchecks who you are

Your namespacejobs, pods, data

1. Install kubectl and the kubelogin plugin

2. Download your config from nrp.ai/config, save it as ~/.kube/config

3. Run any kubectl command: a browser opens for the UCLA login

4. Work in your namespace

On the training JupyterHub this is already set up.

# one-time setupmkdir -p ~/.kubecurl -o ~/.kube/config -fSL https://nrp.ai/config # first command: a browser opens for CILogonkubectl get pods -n my-labNo resources found in my-lab namespace. # make my-lab your default namespacekubectl config set contexts.nautilus.namespace my-lab

https://nrp.ai/documentation/userdocs/start/getting-started/

GPUs

GPUs on NRP

1,411 GPUs in 33 models, contributed by campuses
Most common: A10 and RTX 3090, 24 GB each
About 140 A100s with 80 GB, plus H100s and H200s
Ask for the GPU memory your job needs, not a brand name
What's free right now: nrp.ai/viz/resources

1,411 GPUs across 33 models

A10279

RTX 3090207

A100 80GB SXM118

RTX 2080 Ti106

A100 slices (10 GB)105

L496

GTX 1080 Ti80

V100 32GB68

L4068

RTX A600043

V100 16GB32

RTX A400032

H10024

A100 80GB PCIe24

19 other models129

Up to 16 GB20–24 GB32–48 GB80 GB and up

Choosing a GPU

GPU memory	Examples	Good for	How to ask
Up to 16 GB	RTX 2080 Ti, GTX 1080 Ti, A4000, A100 slices	Teaching, debugging, classic ML, small CNNs	nvidia.com/gpu: 1
20–24 GB	A10, RTX 3090, L4, RTX 4090	Most deep learning; fine-tuning and serving models up to ~8B parameters	nvidia.com/gpu: 1 + memory or model
32–48 GB	L40, V100 32 GB, RTX A6000, A40	Bigger models and batches; serving models up to ~30B	nvidia.com/gpu, or rtxa6000 / a40
80 GB and up	A100 80 GB, H100, H200	Large training runs and large models	A100: access request; H100/H200 reserved

Current PyTorch builds need Turing or newer: on a GTX 1080 Ti or V100, use an older image.

https://nrp.ai/documentation/userdocs/running/gpu-pods/

Ways to Run Work

Interactive: JupyterHub

Pick a GPU from the start menu; explore, debug, teach
Idle servers stop: 1 hour after you close the browser on NRP's hub, 6 hours at most on any hub

Interactive: a pod

Start a container with kubectl, then open a shell in it
Deleted after 6 hours; at most 2 GPUs

Long-running: a Deployment

Keeps a service up (a web app, a database) and restarts it if it crashes
Removed after 2 weeks unless admins grant an exception; no GPU if it sits mostly idle

Batch: a Job

Runs your script to the end, then gives the GPU back
Submit many at once; the rest wait in Pending for a free GPU

$ kubectl apply -f sweep/        # six training jobs

Pendinglr-0.1lr-0.03lr-0.01lr-0.003lr-0.001lr-0.0003GPU 1GPU 2GPU 3GPU 4lr-0.1lr-0.03lr-0.01lr-0.003lr-0.001lr-0.0003time

Six jobs on four GPUs: two wait in Pending and take the first GPUs that free up. When the last one ends, all four GPUs are free again.

1. Pick a GPU

2. Your notebook runs on it

GPUyou work in notebooksidleyou close the browser1 h later: server stops, GPU back to the pool

$ kubectl apply -f pod.yamlpod/dev-pod created$ kubectl exec -it dev-pod -- bashroot@dev-pod:/# nvidia-smi -LGPU 0: NVIDIA A10 (UUID: GPU-…)root@dev-pod:/# python test.py

GPUdev-pod: a shell, tests, debugging6 h: the pod is deleted

Deployment web-app · 2 replicas

web-app-5d59d-x2kqpRunning

web-app-5d59d-m4q9dRunningweb-app-5d59d-m4q9dcrashedweb-app-5d59d-c8w3nRunning

Kubernetes keeps two copies running; when one crashes, it starts a replacement

Nodeweb-appweb-app (replacement)crash: restarted for you2 weeks: removed, unless admins grant an exception

A GPU Job

apiVersion: batch/v1kind: Jobmetadata: {name: gpu-train-daniel, namespace: my-lab}     # (1)spec:  template:    spec:      restartPolicy: Never      containers:      - name: train        image: pytorch/pytorch:2.9.0-cuda13.0-cudnn9-runtime  # (2)        command: ["python", "train.py"]        resources:                                       # (3)          requests: {cpu: "2", memory: 6Gi, nvidia.com/gpu: 1}          limits:   {cpu: "2", memory: 6Gi, nvidia.com/gpu: 1}      affinity:                                          # (4)        nodeAffinity:          requiredDuringSchedulingIgnoredDuringExecution:            nodeSelectorTerms:            - matchExpressions:              - {key: nvidia.com/gpu.memory, operator: Gt, values: ["20000"]}              - {key: nvidia.com/gpu.compute.major, operator: Gt, values: ["7"]}

(1) Where: your namespace

(2) What: a container image, and the command to run

(3) Resources: 2 CPUs, 6 GB of memory, one GPU

(4) Which GPU: more than 20 GB, Ampere or newer; or name exact models

Running the Job

# submitkubectl apply -f gpu-job.yaml # watch: Pending → Running → Completedkubectl get pods -l job-name=gpu-train-daniel # read the outputkubectl logs -f job/gpu-train-daniel # clean upkubectl delete job gpu-train-daniel

Stuck in Pending? kubectl describe pod says why, usually that no matching GPU is free.

Our test: under 30 s from submit to done on an A10 that had the image; several minutes in Pending when the cluster was full.

Output, 9 Nov 2026

Running on: NVIDIA A10Matrix multiply, 4096 x 4096:  CPU (2 cores)  0.09 TFLOP/s  GPU            14.0 TFLOP/s  (154x)Training on 60,000 images, testing on 10,000epoch 1: loss 0.090, accuracy 98.4%, 1.6 sepoch 2: loss 0.009, accuracy 98.7%, 1.0 sepoch 3: loss 0.033, accuracy 98.7%, 1.1 sDone: the job exits and frees the GPU.

Using GPUs Fairly

Run batch work as a Job that exits
Request what you'll use; keep limits within 20% of requests
Keep data on a persistent volume or in S3
Run a sweep as many one-GPU jobs
Borrow idle A100/H100s with priorityClassName: opportunistic, and checkpoint

Don't

Leave a notebook holding a GPU overnight
Put sleep in a Job to keep a GPU (bannable)
Ask for 4 GPUs "just in case": use must stay above 40%
Keep results on the pod's own disk; it vanishes with the pod

https://nrp.ai/documentation/userdocs/start/policies/

AI Services

Hosted LLMs

Open-weight models on NRP GPUs, behind one OpenAI-compatible endpoint:
ellm.nrp-nautilus.io/v1
Free for non-commercial research and education
Prompts are processed on NRP hardware, not sent to a commercial provider
Still not for PHI
About 7.1 billion tokens a day (September 2026)

Model	Good for
qwen3, qwen3-small	Hard reasoning, long documents, images
gpt-oss	General chat, reproducible research
glm-5, kimi, minimax	Agentic coding
deepseek-v4-flash	Whole codebases, very long documents
gemma, gemma-small	Images, video, audio transcription
qwen3-embedding	Search and RAG (embeddings)

https://nrp.ai/documentation/userdocs/ai/llm-managed/models/

Uses from the models page, 26 Oct 2026. The catalog rotates.

Using the LLMs

In the browser: Open WebUI
nrp-openwebui.nrp-nautilus.io
From code: the standard OpenAI library; token from nrp.ai/llmtoken
In JupyterHub: Jupyter AI (in the demo)
Coding agents: Claude Code, opencode, VS Code and others
tested configurations in the NRP docs

Claude Code on NRP's gpt-oss, asked about the demo's GPU job

opencode on NRP's qwen3, same question

Jupyter AI

Already pointed at NRP's models on the training hub, and on any course hub you run
Chat panel: ask about code, an error or a concept
/fix: select a cell that crashed; it explains and proposes a fix
%%ai magics: {variables} in a prompt become their live values, so the model sees your data

%%ai openai-chat:gpt-ossHere is a summary of my dataset: {summary}Which columns would you check first, and why?

Running Your Own Models

Serve your own model: vLLM or SGLang on a GPU in your namespace, your weights or your fine-tune behind the same OpenAI API; ask the admins before running one as a long-lived service
Train and fine-tune: PyTorch and Hugging Face in GPU jobs, from one GPU to eight on a node
Search your documents (RAG): qwen3-embedding and NRP's managed Milvus vector database
Inference accelerators: Qualcomm Cloud AI 100 cards, requested in a job like a GPU

Worked examples: the PEARC26 tutorial (serving, RAG), and a physics neural network trained in GPU jobs (CMS HATS)

https://training.nrp-nautilus.io

Labs and Classes

JupyterHub for Classes

Your class needs	A hub on NRP gives it
The same environment for every student	A container image you choose
No laptop setup on day one	JupyterLab in the browser, with campus login
GPUs for a neural-network unit	GPU options on the start menu, with limits
An AI assistant, without per-seat fees	Jupyter AI and the hosted LLMs, preconfigured
Shared datasets and notebooks	One shared volume in every server

Instructors: contact us; we'll help you set up a hub for your lab or course
It's one configuration file in a namespace you own
Students: if your course needs GPUs, ask your professor to get in touch

Getting Started

1. Log in: nrp.ai (CILogon → UCLA)

2. Request a namespace: get-started.nrp-nautilus.io

PIs, faculty, staff, postdocs; students ask their PI

3. Try the LLMs: nrp-openwebui.nrp-nautilus.io

4. Run your first GPU job:
training.nrp-nautilus.io/ucla-health

Help

Documentation: nrp.ai/documentation
Support chat and email: nrp.ai/contact
Labs, courses or a UCLA campus group: d4diaz@ucsd.edu

Live Demo

Acknowledgments

NRP is supported by NSF awards OAC-1541349, OAC-1826967, OAC-2030508, OAC-1841530, OAC-2005369, OAC-2112167, CISE-1713149, CISE-2100237 and CISE-2120019
Networks: CENIC, Pacific Wave, Internet2, MREN, GPN, NYSERNet, FLR, NEREN, Sun Corridor, OARnet, SCLR and the Albuquerque GigaPoP
The NRP operations teams at SDSC, the University of Nebraska–Lincoln and MGHPCC, and the campuses that host and share their hardware, UCLA among them
Clifford and UCLA Health IT's Research Tech Talks, for hosting

Questions?

Today's materialstraining.nrp-nautilus.io/ucla-health

Portal and loginnrp.ai

Request a namespaceget-started.nrp-nautilus.io

Chat with the LLMsnrp-openwebui.nrp-nautilus.io

LLM tokennrp.ai/llmtoken

GPUs free right nownrp.ai/viz/resources

GPU jobs, docsnrp.ai/documentation/userdocs/running/gpu-pods

Supportnrp.ai/contact

Daniel Diaz · d4diaz@ucsd.edu