GPU Workload Calculator
Size VRAM from model parameters and workload mode, then price the cheapest GPU combination that actually fits.
runs 100% locallyno third-party requestsnot visit-counted
Estimate the VRAM a model needs (inference, LoRA or full fine-tune), see every GPU class that fits and price the hours at the cheapest documented offer - with a sourced centralized reference.
Inputs are sensitive and never leave your browser - this tool skips even the anonymous visit count. Results recalculate as you type.
enter model size…
What the numbers mean
Inference needs roughly (parameters x bytes-per-value) of VRAM for the weights plus activations and KV-cache - the extra factor covers a small context. LoRA fine-tuning keeps base weights plus adapter states; full fine-tuning with Adam needs ~16 bytes per parameter whatever the storage precision. Decentralized markets are strongest for batch jobs and experimentation; for latency-critical production the reliability of a single host matters more than the price. Hourly rates come from the live GPU price index where available, otherwise from our dated rate card.
Questions about GPU sizing
How much VRAM does a model need?
Roughly 16 bytes per parameter for training with optimizer state, far less for inference. The calculator applies those factors and then checks the fit against real GPU classes instead of assuming one card.
Why is the biggest GPU not always the cheapest?
Because cost is per GPU-hour. Nine A100s can undercut four H200s for the same job. The calculator evaluates every class and reports the cheapest workable combination.
Does it use live prices?
It uses the live Akash GPU index where the class is listed and dated constants otherwise, with the cheapest rental SKU shown separately.