GPU memory planner

What it does

A memory budgeting calculator that goes deeper than the usual back-of-envelope formulas. Plug in your config and see a breakdown of weights, activations, KV cache, optimizer state, and gradient accumulation overhead.

What you’ll see

• Per-component memory breakdown (weights, activations, KV, optimizer, gradients)
• Sensitivity sliders: change one variable and watch the others react
• Parallelism comparison: DP vs TP vs PP vs FSDP at the same total budget
• “Will this fit?” check for popular GPUs (A100 40/80, H100 80/96, B200, MI300X)

Status

Beta — formulas validated against published benchmarks for Llama, Mistral, and Qwen families. Mixture-of-experts memory accounting is still being refined.