For lab members using DRAC, please use the lab calendar below to tentatively schedule your job. Please title your event as:
Ensure your job is prepared following the lab’s DRAC parallel job prep guidelines below for improved run efficiency.
For new users of DRAC needing a more detailed explanation, see our detailed DRAC guidelines for beginners.
| #!/bin/bash | |
| #SBATCH --account=def-advisor | user account |
| #SBATCH --gres=gpu:1 ALT: #SBATCH --gpus=nvidia_h100_80gb_hbm3_3g.40gb:1 |
requested GPU. Keep GPU requests aligned with the real parallelism of the model. |
| #SBATCH --cpus-per-task=2 | requested CPUs. Match --cpus-per-task to actual CPU saturation, not to a convenient round number. |
| #SBATCH --mem=64G | memory requested. Match --mem to measured peak memory, not a worst-case guess. |
| #SBATCH --time=3:00:00 | wall time (min 1h, max 72h). shorter given priority. |
Below is a chart outlining the different GPU resources available. DRAC prioritizes jobs based on Reference GPU (RGUs) demand. Fir is the more RGU-intensive cluster. For more information, click here.
CPU and Memory are the recommended maximums per GPU.
| Cluster | | GPU + GPU Memory | | RGU per GPU | | Model/Instance |
(:1 for 1 instance) |
CPU (per GPU) | | Memory (per GPU) | | cli_call |
| Narval | 1/8 GPU, 5GB | 0.57 | gpu:a100_1g.5gb:1 | 1 | 15 | --gpus=a100_1g.5gb:1 |
| Narval | 2/8 GPU, 10GB | 1.14 | gpu:a100_2g.10gb:1 | 3 | 31 | --gpus=a100_2g.10gb:1 |
| Narval | 3/8 GPU, 20GB | 2.0 | gpu:a100_3g.20gb:1 | 6 | 62 | --gpus=a100_3g.20gb:1 |
| Narval | 1 GPU, 40GB | 4.0 | gpu:a100:1 | 12 | 124 | --gpus=a100:1 |
| Fir | 1/8 GPU, 10GB | 1.74 | gpu:h100_1g.10gb:1 | 1 | 35 | --gpus=nvidia_h100_80gb_hbm3_1g.10gb:1 |
| Fir | 2/8 GPU, 20GB | 3.48 | gpu:h100_2g.20gb:1 | 3 | 70 | --gpus=nvidia_h100_80gb_hbm3_2g.20gb:1 |
| Fir | 3/8 GPU, 40GB | 6.1 | gpu:h100_3g.40gb:1 | 6 | 140 | --gpus=nvidia_h100_80gb_hbm3_3g.40gb:1 |
| Fir | 1 GPU, 80GB | 12.2 | gpu:h100:1 | 12 | 280 | --gpus=h100:1 |