gpu-readiness-assessment

GPU Readiness Assessment

Problem

GPU quota (30h/week) is the scarcest resource in Kaggle. It's shared across ALL competitions. Wrong allocation = wasted hours that block other competitions.

Common mistakes:

• Using GPU for tabular competitions where GBDT on CPU is faster and better
• Running AutoGluon with GPU when CPU preset gives same result in less time
• Starting GPU training before verifying the pipeline works on CPU first
• Spending 8h GPU on feature exploration that could be done in 30 min on CPU

The 5-Gate Assessment

Run all 5 gates in order. Stop at the first FAIL.

Gate 1: Task Type Check

Is the task inherently GPU-friendly?
  ✅ Image classification/segmentation (CNN, UNet, ViT)
  ✅ NLP/LLM inference (transformer, BERT, LLaMA)
  ✅ Large-scale deep learning (>1M parameters)
  ✅ RL training (neural network policy)
  ⚠️ Time series with deep learning (LSTM/Transformer — may not beat GBDT)
  ❌ Tabular classification/regression (GBDT dominates, GPU adds nothing)
  ❌ Rule-based agent (PTCG, Orbit Wars — pure logic, no ML)
  ❌ ONNX optimization (NeuroGolf — graph manipulation, no training)

Verdict:
  ✅ → proceed to Gate 2
  ❌ → SKIP GPU, use CPU

Gate 2: Data Size Check

Is the dataset large enough to benefit from GPU?
  >100K rows × >50 features → GPU helps for NN
  >10K images → GPU helps for CNN
  <10K rows → CPU is faster (GPU overhead > speedup)
  <1K images → CPU + transfer learning is sufficient

Kaggle-specific:
  Most Playground series (690K rows) → GBDT on CPU is optimal, GPU NOT needed
  Biohub (3D microscopy) → GPU required (UNet inference)
  ARC-AGI-3 (LLM agent) → GPU required (27B model inference)

Verdict:
  Sufficient data → proceed to Gate 3
  Insufficient data → SKIP GPU

Gate 3: Model Architecture Check

Does the best model for this task require GPU?
  Neural network (CNN/RNN/Transformer) → GPU required for training
  Pre-trained model inference (LLM, UNet) → GPU required for reasonable speed
  GBDT (LGB/XGB/CAT) → CPU only (GPU versions exist but rarely better)
  Linear/logistic regression → CPU trivially
  Rule-based heuristic → CPU trivially

Competition evidence:
  House Prices → GBDT CPU (no GPU needed, 0.11750 LB)
  s6e7 → GBDT CPU (no GPU needed, 0.94942 LB)
  Biohub → UNet GPU (mandatory for 12h inference budget)
  PTCG → Rule-based CPU (no ML training)

Verdict:
  GPU-native model → proceed to Gate 4
  CPU-native model → SKIP GPU