ML Training Cost Estimator

Estimate GPU compute time and cost for training large language models from scratch.

Model Size (parameters)

Dataset Size: 1.0T tokens

1B1T

GPU Model

Number of GPUs

Price per GPU-hour ($)

GPU Utilization: 40%

10%70%

Estimated Results

Training Time

5529575.3 days

132,709,808 GPU-hours

Total Cost (On-Demand)

$3715.87M

8x H100 SXM

With Spot Instances

$1672.14M

Save $2043.73M (~55%)

Cost per GPU-hour

$3.50

per GPU

1Lambda Cloud

$3163.80M

$2.98/hr per GPU

2RunPod

$3344.29M

$3.15/hr per GPU

3CoreWeave

$3524.77M

$3.32/hr per GPU

4AWS (p5)

$4459.05M

$4.2/hr per GPU