Hardware & Compute Guide

Warning

This page is under construction. Specific numbers and benchmarks will be added once hardware configurations are finalised.

Minimum Requirements

Component	Requirement
GPU	TBA — NVIDIA GPU with CUDA support required
GPU VRAM	TBA
System RAM	TBA
Storage	TBA — noise data is large; fast local SSD strongly recommended
CUDA version	TBA
PyTorch version	≥ 2.1 (for `torch.compile` support)

Recommended Configurations

Development (small-scale experiments)

Component	Specification
GPU	TBA
VRAM	TBA
RAM	TBA
Storage	TBA

Production (full O3b run)

Component	Specification
GPU	TBA
VRAM	TBA
RAM	TBA
Storage	TBA

Throughput Estimates

Note

All throughput figures will be filled in after systematic benchmarking. Numbers depend heavily on waveform model, batch size, and torch.compile mode.

Configuration	Batches / second	GPU utilisation
IMRPhenomD, batch 32	TBA	TBA
IMRPhenomPv2, batch 32	TBA	TBA
IMRPhenomPv2 + `torch.compile`	TBA	TBA

Storage Layout

Observing Run data download

data/
└── O3/
    ├── H1/         # Hanford strain (.bin files, one per segment)
    ├── L1/         # Livingston strain
    ├── V1/         # Virgo strain
    └── psds/       # Pre-computed noise PSDs

Estimated total size: TBA (Observing runs are 100s of GB at 2048 Hz sampling rate — download selectively if storage is limited).

Checkpoint storage

Each checkpoint saves model weights, optimiser state, and GradScaler state. Approximate size per checkpoint: TBA.

Cluster Setup

SLURM example

A minimal SLURM batch script will be provided here once the recommended GPU allocation and time limits are benchmarked.

# Placeholder — numbers TBA
#SBATCH --gres=gpu:TBA
#SBATCH --mem=TBAG
#SBATCH --time=TBA
python runs/o3b/train.py

Environment modules

Specific module load commands for common HPC clusters (e.g., CUDA, conda activation) will be listed here once tested.

Tips for Low-VRAM Systems

If your GPU has limited VRAM, try the following in order:

Halve batch_size in RunCFG — roughly halves peak activation memory.
Enable autocast = True in RunCFG to use float16 during the forward pass.
Use ResNet-50 (backend_resnet_size=50) instead of larger backends.
Reduce prefetch in MemmapNoiseSampler.

See also Troubleshooting for GPU out-of-memory fixes.