LLM Scaling Laws & Frontier Simulator D3 Engine

Cross-Entropy Loss Trajectory Loss: 1.92

Compute Allocation Frontier Compute Optimal

Benchmark Emergence Curves

Hardware Allocation & Cluster Costs

H100 SXM5 Cluster-Hours
189.2K
Est. Cost ($3.0/hr)
$567.6K
B200 SXM Cluster-Hours
63.1K
Est. Cost ($4.5/hr)
$283.8K
Assumes 300 TFLOPS/GPU operational efficiency for H100 and 900 TFLOPS/GPU for B200 SXM cluster setups.

Counterfactual Scaling Plateau Laboratory

Exploring Adam Hunt's hypothesis: As token volumes scale past high-quality human data reserves, linear scaling gains collapse due to data degradation and diminishing returns. Move the Plateau slider to simulate data bottleneck caps.

Simulation Execution Summary

Model Parameters: 70.0B
Dataset Volume: 1.40T
Calculated Loss: 1.92
MMLU Score: 68.4%
GSM8K Score: 54.2%
HumanEval Score: 46.1%
Enjoy this tool? Build your own with Super