GPU Computing on Sagehen HPC

Callout

Pomona College HPC Workshop Series

This workshop is part of the Pomona College Research Computing Workshop Series, designed for GPU computing on the Sagehen HPC cluster.

Cluster: sagehen.hpc.pomona.edu Web Portal: OnDemand Support: its-hpc@pomona.edu

Sagehen provides NVIDIA A100, L40S, and RTX PRO 6000 GPUs for deep learning, scientific computing, and data analysis workflows.

Developed for Pomona College by Andrew Wilson, ITS Research Computing. Licensed under CC-BY 4.0.

GPU Computing on Sagehen

Welcome to this workshop! Check the episodes to get started.

Acknowledgments

Developed by Andrew Wilson, Director of Research Computing and Digital Scholarship at Pomona College, with Andrei Motchenko, who tested, edited and produced screenshots for the workshop series.

	Setup Instructions	Download files required for the lesson
Duration: 00h 00m	1. Introduction to GPU Computing	What are GPUs and why use them? When is GPU acceleration beneficial? What workloads suit GPU computing? How do GPUs differ from CPUs? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 00h 20m	2. Sagehen HPC GPU Hardware Configuration	What GPU nodes exist on Sagehen? How do you check GPU availability? What are GPU memory and compute specifications? How do you select specific GPU types? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 00h 50m	3. Requesting and Allocating GPUs	How do you request GPUs in SLURM jobs? What are GPU job parameters? How do you run interactive GPU sessions? How do you allocate multiple GPUs? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 01h 25m	4. GPU Software Setup and Modules	What modules are available for GPU computing? How do you install GPU-enabled software? What CUDA versions exist? How do you manage GPU software environments? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 02h 00m	5. Training Models with PyTorch and TensorFlow	How do you move models to GPU in PyTorch? How does TensorFlow handle GPU automatically? What are best practices for GPU training? How do you optimize GPU memory usage? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 02h 35m	6. Monitoring and Optimization	How do you monitor GPU performance? What metrics matter for GPU workloads? How do you profile GPU code? What optimizations improve performance? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 03h 05m	7. GPU Etiquette and Best Practices	Why is GPU etiquette important? How long can you run GPU jobs? What are fair usage practices? How do you share GPU resources? ::::::::::::::::::::::::::::::::::::::::::::::::
Duration: 03h 30m	Finish

The actual schedule may vary slightly depending on the topics and exercises chosen by the instructor.

Prerequisites

---

Before attending this workshop, you should:

1. Have an active Sagehen HPC account. If you do not yet have one, request access via the HPC account request form (or by emailing its-hpc@pomona.edu). Allow 1–2 business days for account provisioning.

2. Complete the Introduction to HPC Systems workshop (Workshop 0) or have equivalent experience with basic Linux commands, SSH access, and SLURM job submission on Sagehen.

3. Complete the SLURM Job Scheduling workshop (Workshop 9) or be comfortable writing and submitting batch scripts with sbatch.

4. Be familiar with at least one programming language commonly used in GPU computing (Python, C/C++, or CUDA).

Accessing Sagehen

---

You can access the Sagehen cluster in two ways:

Option 1: SSH (Terminal Access)

BASH

ssh <myusername>@sagehen.hpc.pomona.edu

You will be prompted for your Pomona College credentials and DUO Multi-Factor Authentication.

Option 2: OnDemand Portal (Browser Access)

Navigate to https://ondemand.hpc.pomona.edu/ and log in with your Pomona credentials. OnDemand provides a web-based terminal, file manager, and JupyterLab for interactive GPU work.

GPU Resources on Sagehen

---

Sagehen has 10 GPUs across the cluster (confirmed May 2026; see README.md for hardware breakdown) in the gpu partition with the following accelerators:

GPU Model	VRAM	Quantity	Best For
NVIDIA A100	80 GB	4	Large deep learning models, mixed-precision training
NVIDIA L40S	48 GB	4	Inference, visualization, moderate training
NVIDIA RTX PRO 6000	96 GB	2	Large memory workloads, professional rendering, mid-large model training

Callout

Account Limits

Each account/group is limited to a maximum of 4 GPUs simultaneously. Plan your experiments accordingly and release GPU resources when you are done.

Loading GPU Software

---

Sagehen uses the Lmod module system to manage software. To set up a GPU computing environment:

BASH

# Check available CUDA versions
module spider cuda

# Load CUDA toolkit
module load cuda

# Load Python with GPU support (PyTorch, TensorFlow)
module load miniconda3

# Verify GPU access (on a GPU node only)
nvidia-smi

Callout

Important: GPUs Are Only on Compute Nodes

The head node (login node) does not have GPUs. You must request a GPU node via SLURM before running nvidia-smi or any GPU code. Use an interactive session for testing:

BASH

srun --partition=gpu --gres=gpu:1 --time=01:00:00 --pty bash

Pre-Workshop Checklist

---

Before the workshop begins, please verify the following:

• You can SSH into sagehen.hpc.pomona.edu or log into OnDemand
• You can load the CUDA module: module load cuda
• You can load miniconda3: module load miniconda3
• You understand how to submit a basic SLURM batch job

Getting Help

---

If you encounter any issues during setup, contact the HPC team:

• Email: its-hpc@pomona.edu
• Response time: Standard requests are addressed within 1–2 business days