Need help?

The math is the same three-constraint sizing engine the full platform uses (compute / memory / bandwidth, then rounded up to a supported pod topology). Costs use generic public-list cloud rates and lease/CapEx ranges. Real engagements use the customer’s negotiated rates and financial parameters.

Lower precision means fewer bytes per parameter, which reduces both memory footprint and per-token bandwidth. Compute throughput typically scales 2–4× too. We model this conservatively: peak FLOPS multiplier of 2 (FP8) or 4 (FP4), achieved multiplier of 1.8 / 3.2 after dequant and accumulation overhead.

Yes — the calculator is fully usable on the homepage with no signup. Sign up only if you want to size your full fleet, run vendor scenarios, or save plans.

Not yet. The current model assumes inference (per-token throughput targets). Training-specific sizing (gradient memory, optimizer states, activation checkpointing) is on the post-GA roadmap.

NVIDIA H100 / H200 / B200 / GB300, AMD MI300X / MI350X, Google TPU v7 and TPU 8t/8i. The full platform has 20+ chips and updates as new silicon ships.

Email support@mintok.ai with your fleet size (target MW), the workloads you’re trying to price, and any constraints (vendor preferences, compliance needs). We’ll set up a 30-min scoping call.