Get started
Estimator methodology
How compatibility estimates are calculated and ranked.
An estimate, never a benchmark
Weights are estimated from parameter count and numeric precision, then multiplied by a workload allowance: notebook, inference, evaluation, fine-tuning, or full training. The result is divided across the requested GPUs. Runtime memory varies by framework, optimizer, sequence length, batch size, quantization, and parallelism.
Ranking
Compatible machines are ranked by memory fit, final price, observed reliability, network, and environment readiness. Always validate with a small run before committing a large budget.