berth

Adaptive placement infrastructure for AI inference.

Predict what a workload costs and how fast it runs on a given accelerator, before you rent it.

pip install berth-placement
berth place --workload-class voice --model llama3-8b --incumbent h100-pcie --slo-ms 800

Docs · The Placement Index · Defect register

---

The problem

Compute is sold by the hour. It is consumed as work delivered under a deadline. No standard unit connects the two.

So there is nothing to be efficient per. Tokens per second can double while a service delivers nothing sellable, because a token that arrives after its deadline is not a token. Dollars per GPU-hour is the price of the machine, not the cost of the work.

The unit

One output token delivered inside its latency target and above a declared quality floor. A served token.

Everything here measures in that unit. The specification is published and anyone may use it, including people who compete with us. A unit nobody else may use is not a unit.

Three tools

berth predicts what a workload costs and how fast it runs on a given accelerator, before you rent it. Closed-form roofline plus a queueing term. No learned components, roughly a page of arithmetic.

sounding measures the real thing and checks the prediction. It never imports the estimator, enforced by test: evidence that cannot be separated from the thing it evaluates is not evidence.

pilot watches for change, re-decides, and proves what the change saved. It never touches a request.

What we found

Fit the decode constant on one accelerator, then predict a different accelerator it has never seen, on different memory technology.

Worst held-out fold: 9.5 percent, against a 15 percent gate published before the first run.

The residual is the real spread between the two cards. One constant across both costs about nine percent, which is the price of generality and is stated rather than fitted away.

Two more from the same corpus:

Prefill admission is serial, and not as a quirk of one server. Effective parallelism 1.09 under vLLM and 1.01 under SGLang. Modelling that one term drops first-token error from roughly 65 percent to under 9.

Decode is stack-independent to three decimals. 0.854 under SGLang against 0.850 under vLLM, same card, same model.

Every trace is downloadable. Every number is reproducible.

Quick start

from berth import MODELS, WorkloadSpec, profile
from berth.place import decide, render

print(render(decide(workload_class="voice", model_key="llama3-8b",
                    incumbent="h100-pcie", slo_bound_ms=800, batch=8,
                    prompt_tokens=512, output_tokens=128)))

A placement that misses your bound is excluded rather than ranked cheaply. Its cost per compliant token is not high, it is undefined.

Command line

berth estimate	one workload on one accelerator, or all of them
berth place	a decision record: what is recommended, what it beat, by how much
berth premium	rank placements and show the premium for choosing wrong
berth versus	self-host or API, on one axis
berth pilot	one pass of the agent. Shadow by default
berth holdout	check a holdout assignment before opening a period

Adaptive placement

Deciding which chip, which provider, and at which price a model runs under a stated p99, then moving it as prices, capacity, and traffic shift.

Three things make it adaptive.

It is workload-conditional. The answer depends on your token lengths and your concurrency, not on a benchmark average. The right chip can invert inside a single workload as concurrency changes.

It is bounded by your service level. A placement that misses your p99 has no price, not a low one.

It constantly adapts. Prices move, models ship, capacity changes. So does the answer, and most teams decide once and never look again.

pilot closes that loop. It watches model registries, serving-stack releases, provider prices and the corpus; re-estimates when something moves; and opens a pull request against your deployment config with the diff and the evidence attached. It cannot merge, cannot write to a default branch, and cannot touch a path you did not declare.

Most passes produce nothing, and that is the design. An agent that proposes every week gets muted, and a muted agent is worse than none because it looks like coverage.

See the docs.

Proving a saving

A better estimate is not a saving. Proving one needs a declared baseline, a held-out slice of traffic, and an agreed method for comparing them.

That is the Placement Holdout Protocol, published under CC BY so a counterparty can check it before agreeing to be measured.

Contributing a measurement

The corpus grows by measurement. If you run a cell we have not, send it.

Six cells across three memory technologies and two vendors are measured today. Everything else in the fleet is a spec-sheet prior, and every line the CLI prints says which.

When it is wrong

DEFECTS.md is the register: eleven instrument failures, what caused each, how it was caught, and the test that fails if it returns.

Four let bad data through. Seven rejected good data, which is the more dangerous direction: a false negative is caught downstream eventually, and a false positive is silent. One of them nearly discarded the most valuable cell in the corpus.

Only physical impossibility refuses a file. Everything else prints and passes.

A measurement tool that has never been caught lying has not been used hard enough.

---

Apache-2.0. Built by Reckon Research.