The Model Fine-Tuning Throughput Estimator estimates effective training tokens per second and the time required for a planned run. It starts with per-GPU throughput, then applies GPU count, scaling efficiency, and utilization to produce a cluster-level rate.
Use it to compare hardware layouts or reconcile benchmark results with a training schedule. The estimate assumes the workload can scale across the selected GPU count and that input data keeps the devices supplied.
1. Enter a per-GPU benchmark
Use a measurement that matches the model, precision, batch size, and sequence length.
2. Set the cluster size
Enter the number of GPUs assigned to the run.
3. Apply scaling efficiency
Account for communication and synchronization losses.
4. Apply utilization
Reflect stalls, data loading, checkpointing, and other non-productive time.
5. Add total token workload
The calculator converts effective throughput into estimated run duration.
The main result is the estimated sustained token-processing rate for the complete GPU group.
Throughput can change during a run as sequence lengths, batch composition, evaluation, or checkpointing vary.
Given: 1,400 tokens/second per GPU, 4 GPUs, 85% scaling efficiency, 80% utilization, and 108 million training tokens.
Calculation: Ideal throughput = 1,400 × 4 = 5,600 tokens/second. Effective throughput = 5,600 × 0.85 × 0.80 = 3,808 tokens/second. Duration = 108,000,000 ÷ 3,808 = 28,361 seconds = 7.88 hours.
Result: Estimated effective throughput is 3,808 tokens/second and the run takes about 7.88 hours.
Efficiency captures multi-GPU scaling loss; utilization captures time when the available compute is not doing productive training work.
Yes, but a weighted measured average is better than a simple arithmetic average when token volumes differ.
Its effect should be reflected in the benchmarked per-GPU throughput and scaling efficiency.
Input bottlenecks, memory pressure, network contention, evaluation, and checkpoints may not be fully represented.
This calculator is token-based. Convert examples per second by multiplying by average tokens per example.