Estimated results
Results update after calculation.
Scenario comparison
Compare device counts, batch sizes, and precision assumptions.
| Scenario | Devices | Batch | Precision | Training time | Speedup | Estimated cost |
|---|---|---|---|---|---|---|
| Calculate to generate scenarios. | ||||||
Formula used
The selected mode creates the base time before overheads.
Batches per epoch
ceil(training samples ÷ effective processed batch)
Effective batch size
batch size × accumulation steps × devices
Batch-time estimate
epochs × batches × adjusted batch seconds
Throughput estimate
total processed samples ÷ adjusted samples per second
FLOPS estimate
total FLOPs ÷ usable device FLOPs
Complete duration
setup + training + validation + checkpoints + data + overheads
How to use
1. Choose a method
Select a timing method. Use measured values when available. They usually improve estimate quality.
2. Enter configuration
Add dataset and training settings. Include hardware and overhead details. Keep units consistent throughout.
3. Review scenarios
Calculate the estimate. Check warnings and comparison results. Export the report when finished.
Worked example
This example uses one GPU, mixed precision, and measured batch timing.
| Input | Example value | Meaning |
|---|---|---|
| Training samples | 50,000 | Records processed during each epoch |
| Epochs | 20 | Full passes through training data |
| Batch size | 64 | Samples processed per device step |
| Measured batch time | 0.22 seconds | Observed average duration for one batch |
| Validation | 35 seconds | Evaluation time after each epoch |
| Hardware cost | £1.75 per hour | Estimated device rental charge |
Calculation details
A reusable text report appears after calculation.
No calculation has been completed.
Frequently asked questions
How accurate is a training time estimate?
Accuracy depends on input quality. Measured batch time is usually reliable. New workloads can still vary.
Which calculation mode should I choose?
Use batch timing for benchmarks. Use throughput for stable pipelines. Use FLOPS for early planning.
Why is scaling efficiency below one hundred percent?
Devices exchange gradients and parameters. Communication adds waiting time. Data imbalance also reduces scaling.
Does batch size change training time?
Larger batches reduce step counts. Memory limits may block increases. Optimisation behaviour may also change.
How does mixed precision affect duration?
Mixed precision can increase throughput. Compatible accelerators benefit most. Numerical stability should remain monitored.
Are validation and checkpoint times included?
Both are included when entered. Their frequencies control repeated costs. Zero values remove those costs.
Can this calculator track an active run?
Yes, choose progress tracking mode. Enter completed work and elapsed time. Remaining duration is recalculated.
How is training cost estimated?
Duration multiplies hourly device cost. Device count is included. Electricity cost appears separately.
Why might the real finish time differ?
Thermal throttling can occur. Storage and network contention can change. Interruptions also extend real training.