ferrotherm-cloud

Real fabricated Ising silicon, reached through the same ferrotherm::fabric::Device trait as a CPU.

Currently one fabric: Hitachi's CMOS annealing machine, through Annealing Cloud Web. A 384×384 King's graph with four-bit coefficients, and a free public API that essentially nobody has used — two papers in all of OpenAlex mention the service. It is therefore the cheapest real fabric in the world to support, and supporting it is what makes "runs on any fabric" checkable rather than rhetorical.

You bring your own credentials

This crate ships no token, no account and no default identity. It talks to an account you create yourself, and until you do, it does nothing — there is exactly one network call in the crate and reaching it takes a token you supplied, a program you laid out, and a call you made.

Getting a token

1. Request one at https://annealing-cloud.com/en/web-api/token-request.html. The form asks for an email address and a country, and requires agreeing to two conditions: that you will not use the site or its output data for any purpose including the development of weapons of mass destruction (their Terms of Use, Section 8, Export Controls), and that you consent to the collection of personal information under those Terms. Intended use and user type are optional.
2. The administrator emails the token back to you. That page does not state how long issuance takes or what the usage limits are; the service homepage describes the Web API as free.
3. Put it in your environment, never in a file you commit:

export ACW_TOKEN=<the token they emailed you>
cargo run --release -p ferrotherm-cloud --example hitachi_run

Hitachi::from_env reads exactly that one variable and returns Err when it is unset. It does not fall back to a bundled key, a config file or a credential store, because there are none. Hitachi::new takes the token directly if you would rather source it another way.

Read the API you are calling first: https://annealing-cloud.com/en/web-api/reference/v2.html.

Where it sends

The default endpoint is ACW_ENDPOINT — the public solve endpoint of the service this driver is named for. It is a default, not a fixture:

let d = Hitachi::new(token, Machine::Asic).with_endpoint("http://127.0.0.1:8080/solve");

endpoint() tells you where an instance will post. A test in this crate pins the guarantee: with ACW_TOKEN unset there is no device at all, and pointed at a port nothing can listen on, describing the fabric and laying a program out still succeed — because those paths are local.

What it does before it sends anything

Everything except the submission itself is local, and most of the value is there:

• Sign conventions, measured rather than assumed. Their energy is Σ pᵢⱼ sᵢsⱼ, minimised, so a positive coefficient is antiferromagnetic — the opposite of ferrotherm's -Σ Jᵢⱼ sᵢsⱼ. That was established empirically on the first call, not read off a document: four positive couplings on a 2×2 block came back as a checkerboard at energy −4. A sign error here produces entirely plausible output that is wrong on every problem.
• Layout and refusal. Sites are grid coordinates; neighbours are the eight surrounding cells, orthogonal and diagonal. A coupling between non-adjacent coordinates is an error, not a silently dropped term. place() embeds a model that does not already fit; layout() requires one that does.
• Limits declared before submission. The ASIC stores coefficients in four bits, -7 ≤ p ≤ 7, integers — so J = 0.5 is refused here rather than quietly rounded on the way in. Knowing that before you submit is the difference between a refused job and a wrong answer.

cargo add ferrotherm-cloud

Apache-2.0. Part of ferrotherm, from the Institute for Physical AI @ BMI.