Training Cost Calculator

A training cost calculator estimates the total compute expense of training a machine learning model from scratch or fine-tuning an existing one. Unlike inference costs — which scale with requests — training costs are a one-time capital expense driven by three factors: model size (parameters), dataset size (tokens), and GPU cluster cost. The industry-standard approach uses the Chinchilla scaling law formula: FLOPs ≈ 6 × N × D, where N is the number of model parameters and D is the number of training tokens. Our training cost calculator applies this formula together with real-world Model FLOP Utilisation (MFU) to produce wall-clock time and total cost estimates that match production training runs.

1. Configure Model and Dataset: Enter the model size in billions of parameters, the dataset size in billions of tokens, and the number of training epochs. The calculator multiplies dataset tokens by epochs to get the total effective token count — the actual number of tokens the model sees during training.
2. Select Your GPU Cluster: Choose from 12+ cloud instances across AWS, GCP, Azure, Lambda Labs, RunPod, and Vast.ai, or enter a custom price. Set the total number of GPUs in your cluster and adjust MFU (Model FLOP Utilisation) to reflect real-world hardware efficiency — typically 35–50% for well-optimised LLM training.
3. Choose Training Type: Full fine-tuning trains all model parameters. LoRA and QLoRA update only a small fraction (0.1–5%) of parameters, dramatically reducing compute and memory requirements. The training cost calculator adjusts the FLOP formula for each training type.
4. Review Cost and Compare Providers: The results panel shows total training cost, wall-clock time, tokens per second, cost per billion tokens, and a ranked comparison of all supported cloud instances for the same workload.

• FLOPs (6ND Formula): The Chinchilla paper established that a full forward + backward pass through a transformer requires approximately 6 × N × D floating point operations, where N is parameters and D is tokens. This is the foundation of every serious training cost calculator.
• Model FLOP Utilisation (MFU): No GPU cluster achieves its theoretical peak TFLOPS in practice. Communication overhead (all-reduce between nodes), memory bandwidth stalls, and pipeline bubbles reduce effective compute to 35–55% of peak. MFU is the single most important tuning variable in an accurate training cost calculator.
• Spot vs. On-Demand Pricing: AWS, GCP, and Azure offer preemptible/spot instances at 55–70% discounts. For long training runs with frequent checkpointing, spot instances can reduce training costs by more than half — at the cost of potential interruptions.
• LoRA / QLoRA FLOP Reduction: Parameter-efficient fine-tuning methods like LoRA and QLoRA only compute gradients for a small set of adapter parameters (typically 0.1–5% of total parameters), significantly reducing the backward-pass FLOP count and total training cost.

The training cost calculator runs entirely client-side in your web browser. Your model configurations, dataset sizes, cost inputs, and all calculated projections are processed locally on your device and are never transmitted to any server. No sign-up or account is required, and no data is retained between sessions. The tool is 100% free with no usage limits — your training budget planning data stays completely private.