GPU Selection

Know before you run.

Analyze your PyTorch code to predict memory and runtime across GPU types.

GPU Analysis
Complete
H100
4h 12m · $13.85
<1% OOM
2
A100 80GB
6h 31m · $12.97
8% OOM
L40s
Insufficient VRAM
>99% OOM
Recommended H100 (best cost/perf)
// Predictions

What we analyze.

Memory usage

Model weights, gradients, activations, optimizer state, and mixed-precision overhead.

Runtime

Training or inference loop time based on estimated iteration time per GPU type.

OOM probability

Likelihood of out-of-memory errors based on peak memory vs. available VRAM.

Cost estimate

Predicted cost based on runtime and GPU pricing. Find the best cost/performance ratio.

// How it works

Static analysis, not execution.

1
Submit
Point at your Python script. We trace the compute graph without running it.
2
Analyze
AI predicts memory and runtime for each GPU based on your model architecture.
3
Choose
Get ranked GPU recommendations. Pick manually or let Lyceum auto-select.
PyTorch PyTorch Lightning
// Scope

What works today.

Supported

  • Single Python script entry point
  • Single model training or inference
  • Imported models (HuggingFace, torchvision, timm)
  • Command-line arguments

Not yet supported

  • Jupyter notebooks
  • Multi-GPU / distributed training
  • Custom CUDA extensions
  • TensorFlow, JAX
// Usage

One command.

Terminal
lyceum gpu-selection run train.py

Pass script arguments after --: lyceum gpu-selection run train.py -- --epochs 10

Stop guessing.

GPU selection is free for all Lyceum accounts.