Supported silicon backends

This page is the public entry point for accelerator and chip teams searching for fak backend conformance, fak-certified backend, neo-silicon agent kernel, bring-your-accelerator agent serving, or vendor-neutral inference backend.

The short version: fak has a binding layer for model execution. A new accelerator does not fork the agent kernel; it implements internal/compute.Backend, registers a backend name, advertises exact Caps, and proves its correctness class against the reference floor.

Start here: bring a backend and prove it runs

Audience: an accelerator or chip team bringing a backend to fak’s kernel — you implement internal/compute.Backend, register a name, advertise Caps, and prove a correctness class against the cpu-ref floor. fak keeps the agent-kernel boundary stable; you own the kernels and the runtime evidence.

Next action: prove your backend actually executes instead of silently falling back to the CPU reference — the non-reference gate fails closed if it does:

cmd/modelbench -backend <name> -require-non-reference

Then label the support row by the evidence you have (Reference / Approx / FENCED / PARTIAL / not-yet, defined below). Honest fence: fak-certified backend is the intended public mark but is not-yet — do not claim it until the Backend Conformance Kit lands.

What is shipped

Surface	Status
internal/compute.Backend	Shipped whole-op interface for model execution.
Caps	Shipped optional capability vocabulary: async, fused attention, graph compile, dtype upload, device memory, collectives, and capacity probes.
cpu-ref	Shipped Reference backend and correctness floor.
Device backends	CUDA and Vulkan are documented with hardware witnesses in the HAL explainer.
Non-reference gate	cmd/modelbench -backend <name> -require-non-reference fails closed when a run silently falls back to CPU.

The detailed contributor-facing seam is documented in hardware portability via the compute HAL.

What is not yet a mark

fak-certified backend is the intended public mark, but it should not be used as a claim until the Backend Conformance Kit lands. Until then, support rows should use plain evidence labels:

Label	Meaning
Reference	Exact-reference rungs are green.
Approx	Approximate parity gates, such as argmax and logit cosine, are the right witness.
FENCED	The backend has a named seam, cap, or support row that bounds the claim.
PARTIAL	Some methods, model shapes, or caps work, and the gaps are listed.
not-yet	The path is designed or planned but lacks a witness.

Vendor-neutral support row

A backend support row should name the facts a router and a customer can actually use:

Field	Why it matters
Backend id	Stable selector used by the registry and by route logs.
Correctness class	Prevents an approximate accelerator from being judged as byte-identical reference.
Caps	Says what the backend implements, not what the device family can do in theory.
Hardware and runtime	Makes measurements reproducible.
Model shape	Keeps one model’s witness from becoming a universal claim.
Capacity	Real bytes only when CapacityProbe or HostCapacityProbe is advertised.
Parity witness	Shows the backend was compared to the cpu-ref floor.
Fallback	Names when and where a route falls back.

Why this matters for agent serving

Agent serving is not only throughput. A backend sits under policy, audit, tool-call admission, context, cache provenance, and refusal reasons. The value of the seam is that a silicon vendor can focus on kernels and runtime evidence while fak keeps the agent-kernel boundary stable.

For cloud operators building heterogeneous fleets, see heterogeneous silicon fleets.