Pareto explorer — the idea, on an example

A sizing tool that returns one "best" configuration has quietly discarded hundreds of other valid ones — and with them every trade-off you might have preferred. The Pareto explorer (the Pareto explorer results view on the TCO page) keeps them all: every mark is a complete serving setup — GPU model, TP/PP/EP topology, batch size, replica count — that meets the same SLA. Fleet cost sits on one axis, per-token latency (TPOT) on the other; lower is better on both.

The idea: dominated setups and the front

A configuration is dominated when another one is both cheaper and faster — no rational buyer picks it, whatever the budget. Everything not dominated forms the Pareto front: the staircase in the chart. On the front you cannot improve one axis without paying on the other, which is exactly what makes it the shortlist worth debating.

The front is computed on fleet $/hr against TPOT, both minimized. Prices come from the same live market medians as everywhere else on the site, so the staircase moves with the market — re-run the sweep on decision day.

Reading it, on an example

• The ends of the staircase are the headline options. The cheap end (16× A100 at $22.4/hr · 49 ms) is the cheapest way to hold the SLA; the fast end (6× B200 at $38.4/hr · 26 ms) is the fastest. Everything between is a legitimate middle ground.
• Steps price your latency. Moving from 8× H200 ($31.8/hr · 34 ms) to 6× B200 ($38.4/hr · 26 ms) buys 8 ms for +$6.6/hr — about $4,800 a month. Whether that is worth it is a product decision, and now it is a visible one.
• Grey marks are the argument for the chart. The 10× H100 NVL configuration ($27.9/hr · 44 ms) looks reasonable in isolation — but the front holds one that is cheaper and faster (10× H100 SXM at $24.9/hr · 42 ms), so it is dominated. A single-answer tool might still have offered it.
• Dashed rings mean a wider error bar. Pipeline-parallel, TP above 8 and PCIe or xGMI multi-GPU setups (the 40× L40S point) carry a less calibrated throughput estimate — treat their position as approximate.
• Configurations without a market or catalog price cannot be plotted — they have no cost coordinate — so the legend counts them instead of dropping them silently.

Using it in the tool

Open the planner with your model, users and TPOT target, then switch Results view from Best per GPU to Pareto explorer. The page sweeps every feasible configuration — GPU × tensor, pipeline and expert parallel shape × batch — at your SLA. Hover a mark to see its unit; click it (tap on a phone) to open the detail: topology, total GPUs, fleet $/hr and $/1M tokens, with a link to the GPU page. The explorer works in chat mode only; agentic plans use the options table.

The Optimize map asks a related but different question: it plots every setup that fits for one model by speed per user and $/1M tokens, without an SLA or a user count. Use Optimize to understand the space, the Pareto explorer to price a fleet against a target. See Reading the Optimize map.