berth is a placement estimator for inference engineers

The same workload can cost multiples more on one chip than another. And the premium is not a constant: it moves with the model, the context length, the batch size, and the latency target. That variance is the reason berth exists. Not a lookup table, but an estimator you run per workload, across every chip and provider you could place it on.

Here is one placement we measured end to end. For Llama-3-8B decode, an L40S is 1.6x cheaper than an H100 PCIe at matched concurrency. The H100 is faster, but if your p99 target does not require that speed, you are paying 1.6x for headroom you will never use.

Every estimate is a closed-form roofline you can audit by hand. Decode is a race between compute and memory bandwidth. Prefill is compute plus a fixed scheduling floor. Concurrency is serial batch admission. No learned model stands between you and the number, and every prediction ships with the term that produced it.

On real silicon it holds: decode latency predicted to 4.2 percent on an H100 PCIe and 10.6 percent on an L40S, using spec-sheet priors with no tuning, against a 15 percent gate published before the first run.

You describe a workload; berth returns dollars per million tokens, time to first token, time per output token, and the placement premium: how much you would save on the cheapest silicon that still meets your p99.

berth is an estimator. It predicts from a model, and an estimate you cannot check is just an assertion. So berth ships sounding, the same measurement harness that produced its own validation data. You use the estimator; the harness is how we earned your trust in it.

sounding drives an OpenAI-compatible completion server, vLLM or SGLang, with streaming requests across a grid of batch sizes, prompt lengths and output lengths, and records time to first token and time per output token per cell. Standard-library Python, so the box you rent needs nothing beyond a Python interpreter and a running model server.

One run measures one accelerator against one model: ninety cells across five batch sizes, three prompt lengths and two output lengths, in randomized order so thermal drift does not confound the sweep. Roughly an hour on a rented box, a few dollars of GPU.

Add a card by running it again there and passing both files to validate. Cells accumulate, and that is how the L40S and H100 PCIe comparison above was built.

The fixed prefill floor is a property of one accelerator, driver, server and config, and is not predictable from a spec sheet: it measured 74.6 ms on the L40S and 54.6 ms on the H100 PCIe. Silicon profiles ship 0.0 and every run fits its own. Do not carry someone else's.

No GPU handy? --mock runs the whole path from sweep to error report against the analytical model plus noise. Mock traces are stamped as such and can never enter the corpus.

Profile one real workload. If berth's number is useful, run a sounding and send the anonymized trace so the public estimate gets more accurate for everyone serving that workload shape. Disputes that arrive with traces attached outrank everything else.

The corpus is worth exactly as much as the provenance behind it.

Compute is bought by the GPU-hour and consumed as work under a deadline. No standard unit bridges the two, so nobody can say what a token costs, which is why depreciation schedules blur, why circular financing is unfalsifiable, and why the bubble argument compares capital expenditure in dollars against demand in vibes.

Every general-purpose infrastructure has crossed this point, and each one resolved it the same way. Railways got a clearing house. Power got a kilowatt-hour. Freight got a box. Money got interchange. Packets got a protocol. In each case the physical layer was necessary, competitive and margin-poor, and the layer that measured and routed was singular, standard and margin-rich.

berth is the instrument. What it measures is the unit.