Claude Haiku 4.5
Estimating GPT-3 Training FLOPs
I'll work through this from first principles using publicly known GPT-3 specs.
Known Parameters
- Model size: 175 billion parameters
- Training tokens: ~300 billion tokens (estimated from OpenAI's hints about compute)
- Training approach: Standard transformer with dense attention
Step 1: FLOPs per Token per Parameter
For a forward pass through a transformer:
- Each token attends to all previous tokens (roughly)
- Each parameter is involved in roughly 2 operations per token (multiply-accumulate = 2 FLOPs)
Rough estimate: ~2 FLOPs per parameter per token for forward pass
Step 2: Account for Backward Pass
Standard backpropagation requires:
- Forward pass: 1x
- Backward pass: ~2x (gradients for weights and activations)
Total: ~6 FLOPs per parameter per token (accounting for forward + backward)
Some literature suggests 3-6x multiplier; I'll use 6 as a reasonable middle-ground estimate.
Step 3: Calculate Total
Total FLOPs = Parameters × Tokens × FLOPs-per-token-per-param
= 175 × 10^9 × 300 × 10^9 × 6
= 175 × 300 × 6 × 10^18
= 315,000 × 10^18
≈ 3.15 × 10^23 FLOPs
Final Estimate
~3 × 10^23 FLOPs (order of magnitude: 300 exaFLOPs-seconds)
This aligns with OpenAI's reported ~3,640 petaflop/s-days, which converts to roughly 3 × 10^23 FLOPs.
Reasoning Check
- This represents ~10,000 GPU-years of A100 compute (rough validation)
- Cost would be $10-15M at 2020 cloud prices
- Publicly consistent with OpenAI's disclosed estimates
Confidence: Moderate — the 2-6x multiplier for backprop is the biggest uncertainty, but order of magnitude should be solid.
