LLM Scaling Tradeoff Simulator D3.js Lab

Explore Pre-Training Scaling Laws vs. Test-Time Search Compute Allocations

Presets:
Compute & Model Specs
10^25 FLOPs
70 B
2.0 T
32 steps
8 branches
Compute Allocation Pareto Frontier
Pareto Optimal
Current Operating Point
Predicted Task Accuracy
84.2%
Optimal balance achieved
Inference Latency
1420 ms
MCTS & Token generation
Cost / 1M Queries
$18.40
FLOPs converted to USD
Detailed Compute Distribution & Breakdown
Compute Domain Parameter / Metric Allocated FLOPs Share of Total Compute Operational Impact
Pre-Training Scaling 70B Params / 2.0T Tokens 8.40e+23 65.0% Establishes baseline knowledge & parametric memory
Test-Time Search 32 Search Depth × 8 Branches 4.52e+23 35.0% Drives MCTS reasoning expansion & error correction
Total Pipeline Budget System Operating Point 1.00e+25 100.0% Combined amortized FLOP utilization

Scenario Summary & Optimal Allocation Proof Target

Target Accuracy 0.8420
Optimal Pre-train FLOPs % 65%
Optimal Inference FLOPs % 35%
Inference Latency (ms) 1420
Cost / 1M Queries (USD) 18.40
Enjoy this tool? Build your own with Super