Euler GPU Track

A GPU is an accelerator that helps only software written to use it. On Euler, you request a GPU as part of a scheduled job; the request reserves shared hardware even when the program does not keep that GPU busy.

Complete this track before regular GPU use on Euler. GPU resources are available only through shareholder accounts such as es_fuge; the public share has no GPUs.

This page summarizes the track. Follow the current mission in your local Passport; do not use the links below as a second checklist.

Prerequisites#

• Passed Euler CPU Track
• Project code has a real GPU implementation
• Supervisor confirmation that GPU use is appropriate

Current Training Evidence#

The initial mission checks one synthetic script profile and choices. It requires no GPU allocation and demonstrates neither performance nor a live scaling measurement.

Use the current Passport lesson for sequencing. The competencies and evidence below describe later project capability and private owner review; they are not extra public submissions or automatic access grants.

Competencies#

You can:

• justify why a workload needs a GPU;
• request one RTX 4090 by default or one RTX 3090 as a general fallback through es_fuge;
• recognize when the two RTX PRO 6000 GPUs are appropriate and verify CUDA 13 compatibility before requesting pro_6000;
• select CPU, system memory, GPU count, and time intentionally;
• perform the smallest useful GPU smoke test;
• inspect GPU utilization and memory with nvidia-smi;
• recognize CPU/data-loader and I/O bottlenecks;
• checkpoint work before the wall-time limit;
• use euler-tunnel for short interactive debugging and sbatch for training;
• establish a one-GPU baseline before requesting multiple GPUs;
• cap GPU arrays so they cannot consume the full shared allowance accidentally.

Mission Coverage#

1. Read IDEAL Lab Euler share policy.
2. Read Slurm GPU jobs.
3. Read Euler tunnel.
4. Complete GPU resource review.

Evidence#

• A reviewed one-GPU Slurm script using es_fuge.
• A reasoned RTX 4090 default, RTX 3090 fallback, and explanation of why RTX PRO 6000 is or is not required.
• A monitoring plan for GPU utilization, GPU memory, CPU, and system memory.
• A checkpoint/recovery plan for training jobs.
• A multi-GPU scaling criterion, even when only one GPU is currently requested.

A live GPU job is not required for the initial review exercise. If the project requires a smoke test, keep it short and obtain supervisor approval.

Understand Before Accepting AI Output#

• The program executes GPU operations rather than only detecting CUDA.
• I can explain every requested CPU, memory, GPU, and time value.
• I know why adding another GPU may not improve a CPU-bound pipeline.
• I will move long or unattended work from a tunnel to sbatch.

Pass Criteria#

The student corrects unsafe GPU scenarios and explains a measurement-driven plan. Multi-GPU access is not justified solely by dataset or model size.