Will it fit — and how fast?
Pick a model size, quantization, and node count. Estimates use per-token costs measured on our 8× GB10 cluster — including the multi-node overhead the spec-sheet math ignores — and the result box says how far our measured results have landed from them.
—
Enter your model above.
This is a rough estimate, not a floor or a promise: the time to read the active weights from memory, plus fixed per-token costs we measured on our older builds. The pure bandwidth limit (shown separately) is a ceiling for decoding without speculative decoding; speculative decoding can beat it because one weight sweep can yield several tokens. How close is the estimate? Across our measured entries, results without speculative decoding landed between about 0.5× (untuned, no CUDA graphs) and 3× (our newest stack) of it; with speculative decoding, up to about 6×. That is too wide for a ± figure — treat it as an order-of-magnitude sanity check. Multi-node figures assume a correctly configured 100 Gbps+ fabric; a misconfigured one costs more than any number here (see the RDMA ceiling note).
The calculator says one thing. Your cluster says another?
That gap is exactly what we fix. The 1-day diagnostic measures where your real number is, why, and what's recoverable — in writing, or it's free.
Book a 1-day diagnostic →