Use this calculator to estimate useful throughput for a fine-tuning run. Adjust the workload and operating assumptions to match a pilot run or planned deployment. The result is a planning estimate, not a guaranteed hardware benchmark or provider quote, and the supporting metrics make the main estimate easier to audit.
The main result reports useful throughput under the assumptions entered. Use it to compare configurations, budgets, or delivery targets on a consistent basis.
Actual results can vary with model architecture, sequence length, batching, hardware, software stack, queueing, failures, and validation policy.
250 million tokens over 8 hours equals 8,680.56 tokens/s. At 100% useful output across 8 GPUs, that is 1,085.07 tokens/s/GPU.
No. It uses the operational values you enter, so you can model your own hardware, provider, and workload.
Use measurements from a representative pilot when possible. Peak specifications are less reliable than sustained observed values.
Yes, fields that naturally support fractions accept decimals. Counts that represent deployable GPUs are rounded up in the result where applicable.
Only the overhead represented by the visible inputs is included. Queueing, data transfer, failures, and human review may need a separate allowance.
Model size, sequence length, batching, hardware, software, rate limits, validation rules, and utilization can all change actual performance.
| Item | Meaning |
|---|---|
| Main result | Useful throughput |
| Workload | Total work represented by the entered values |
| Capacity assumptions | Sustained values, not guaranteed peak specifications |