Model Training Cost Calculator
Estimate GPU-hours and dollar cost to train a model from parameters, tokens, and GPU type, with a cloud provider price comparison table. Free and client-side.
Loading tool...
What is Model Training Cost Calculator?
The Model Training Cost Calculator applies the standard transformer compute rule - about six FLOPs per parameter per token - to your parameter and token counts, then converts total compute into GPU-hours at your GPUs peak TFLOPS and utilization, and into dollars at your hourly rate, with a four-provider cloud price comparison table.
Key Benefits
- Budget a training run before you reserve a single GPU
- The six-N-D rule gives defensible first estimates instantly
- Utilization input keeps estimates honest about your cluster
- Provider table turns GPU-hours into concrete dollar options
- No account, no API key - everything computed in the browser
Common Use Cases
- •Scoping a pretraining project for leadership budget approval
- •Comparing cloud providers before reserving GPU capacity
- •Checking whether a Chinchilla-optimal token count fits the budget
- •Teaching ML teams why utilization dominates training bills
How to Estimate Training Cost
- Enter model size and tokens: Type the parameter count in billions and planned training tokens - compute FLOPs appear instantly.
- Pick the GPU and utilization: Choose your GPU type and set a realistic utilization so GPU-hours reflect your cluster.
- Compare providers: Enter your own hourly rate, then scan the provider table for concrete budget options.
Key Features
- Chinchilla-style compute estimate from parameter and token counts
- GPU presets for A100, H100, and L40S with peak TFLOPS
- Editable utilization so effective FLOPS match your real cluster
- GPU-hours and dollar cost at your own hourly rate
- Four-provider cloud price comparison table per GPU
- Pure client-side math - no account, no API key