GPU Workload Calculator - VRAM & Hourly Cost
Estimate the VRAM a model needs, find the cheapest GPU class that fits, and price the hours - with an AWS/Lambda reference so you can see what "cheaper than the cloud" actually means.
enter model size…
The 2-bytes-per-parameter rule
Weights need roughly (parameters x bytes-per-value) of VRAM; the KV-cache and activations add more, especially at long context. For inference you can often fit a 7B model on 24 GB and a 70B model needs either multi-GPU or heavy quantisation. Decentralized markets are strongest for batch jobs and experimentation; for latency-critical production the reliability of a single host matters more than the price.