Describe the data
Total tokens (or examples x average length) for LLMs; images and resolution for vision. Epochs.
Fine-tuning
Training Time & Cost Estimator
Given tokens or images, epochs and a realistic throughput for the GPU, this gives wall-clock time and cost at a per-hour price, with the throughput presets measured on common cards. It also works backwards: given a budget, how many epochs or examples fit.
Loading the tool…
How it works
Pick the hardware
Throughput presets for T4, L4, A10G, A100, H100 and consumer cards, or type your own tokens per second or images per second.
Read time and cost
Hours, and cost at the price you enter (presets for common providers); a budget mode inverts the calculation.
Formats, limits and the numbers that matter
Throughput reality
Presets are for LoRA at sequence length 2048 with bf16 and gradient checkpointing; full fine-tuning is 2-3x slower, QLoRA 1.5-2x slower than LoRA.
Overheads
Add evaluation passes, checkpoint saves and startup; the estimator adds 10 percent by default.
Vision
Images per second scales with resolution squared; the preset numbers are at 640 for YOLO and 224 for classifiers.
Common questions
Tokens or examples?
Enter whichever you know; the Tokenizer Playground or LLM Dataset Builder gives the token total for a file.
Is spot pricing included?
Enter the spot price; the tool does not model interruptions.
How accurate?
Within about 30 percent if throughput matches; measure 50 steps on the real machine and paste the rate for a tight number.