Enter model training inputs
Use the calculator to estimate full project cost, not only compute. The form uses a responsive grid with three columns on large screens.
Example data table
Use these sample values to test the calculator and compare small, medium, and large model training scenarios. Estimated totals are illustrative.
| Scenario | GPUs | Hours per run | Runs | GPU rate | Storage GB | Engineer hours | Estimated total |
|---|---|---|---|---|---|---|---|
| Prototype fine tune | 4 | 18 | 2 | $2.20 | 800 | 10 | $1,199.03 |
| Mid scale retraining | 8 | 72 | 3 | $3.20 | 2500 | 28 | $10,425.37 |
| Large experiment sweep | 16 | 120 | 5 | $4.60 | 6000 | 70 | $70,984.33 |
Formula used
The calculator combines direct and indirect spending to estimate full training budget.
- Effective runs =
planned runs × (1 + retry overhead ÷ 100) - Baseline GPU hours =
GPU count × hours per run × effective runs - Billed GPU hours =
baseline GPU hours × (1 + orchestration overhead ÷ 100) - Gross compute cost =
billed GPU hours × GPU hourly rate - Net compute cost =
gross compute cost − discount - Energy cost =
energy kWh per hour × hours per run × effective runs × energy rate - Storage cost =
storage GB × retention months × storage rate - Labor cost =
engineer hours × engineer hourly rate - Transfer cost =
data transfer GB × transfer rate - Subtotal =
net compute + energy + storage + labor + transfer + dataset prep + misc fixed - Total cost =
subtotal + (subtotal × contingency ÷ 100)
How to use this calculator
- Choose the currency that matches your budget planning.
- Enter compute inputs, including GPU count, hourly rate, and wall clock hours per run.
- Add retry and orchestration overhead to reflect failed jobs, scheduling delays, checkpoints, and cluster management load.
- Fill in supporting costs for energy, storage, engineering time, data transfer, dataset preparation, and other fixed expenses.
- Apply any reserved or spot discount, then add contingency to cover uncertainty.
- Press Calculate training cost to display the result above the form.
- Use the CSV and PDF buttons to export the cost summary after calculation.
Frequently asked questions
1. What makes this calculator advanced?
It estimates more than GPU time. It also includes retries, orchestration overhead, storage retention, power usage, labor, transfer fees, fixed project costs, discounts, and contingency.
2. Should I use planned runs or total attempts?
Enter the number of runs you intend to complete. The retry overhead field automatically increases total attempts, helping you budget for failed jobs and reruns.
3. What does orchestration overhead mean?
It represents extra billed cluster time from setup, checkpoint handling, distributed coordination, logging, and idle transitions that extend paid runtime beyond core training work.
4. Why include energy cost if I already pay cloud rates?
Include energy only when it is billed separately or when you are planning on premises infrastructure costs. For bundled cloud pricing, you can set energy inputs to zero.
5. Can I use it for fine tuning and full pretraining?
Yes. The calculator is workload agnostic. As long as your inputs reflect the expected hardware, runtime, storage, and support effort, the budget estimate remains useful.
6. How do I estimate the discount field?
Use the percentage savings from reserved capacity, spot usage, long term commitments, or negotiated pricing. Leave it at zero if your environment uses on demand rates.
7. What should contingency cover?
Contingency can cover scope growth, extra validation runs, longer retention, debugging time, data expansion, model changes, or price movements during the training window.
8. Can I export the results for reports?
Yes. After calculation, use the CSV button for spreadsheet analysis or the PDF button for a clean shareable summary of the cost breakdown.