Hardware

What is documented here

Wirken runs on commodity hardware. The four configurations below are reference builds for operators who want a hardware-side answer to "what should I run wirken on?" None is a product line. All four are parts lists. You or your supplier can buy the same parts and build the same machine.

Spule

Spule is wirken on a Raspberry Pi 5: switchboard, adapters, audit log, and WebChat on the local box, LLM over TLS, no inference weights held locally. Wirken is a single static binary linked against musl, and the aarch64 build installs from the same one-line installer with no runtime underneath it. The Cortex-A76 includes the ARMv8 cryptography extensions, so SHA-256 hashing for the audit chain uses hardware instructions. Spule has less capacity than Fern for concurrent channels and retained history.

These parts ship as a single kit, the BerryBase Raspberry Pi 5 NVMe Starterkit, at about €229.

Fern

Fern hosts the switchboard, the adapters, the audit log, and the WebChat surface, and reaches the LLM over TLS. The local box never holds inference weights. Fern fits the confidential-inference posture cleanly: the only outbound inference connection is an attested HTTPS path, which can be pinned at the network layer.

Heim

Heim is the aggregation box. It runs inference on the machine itself: embedding generation and the librarian side of Zirkel, wirken's research aggregator, and the scoring passes of Lyrik, the static-analysis skill. There is no outbound LLM connection.

RAIV

RAIV, a redundant array of inexpensive videocards, is an inference server for open-weight models rather than a wirken host. Seven used 32 GB datacenter GPUs form two mirrored groups of three plus a spare. One group holds a 100B-class model at four-bit with room for context. Each group runs its own llama-server process, and nginx or Caddy in front checks /health on both and fails over: when a card in one group fails, the other group serves, the spare replaces the failed card, and one in-flight conversation is lost. Wirken, on a Fern- or Heim-class host, reaches it over the network as an OpenAI-compatible endpoint. Recent Ollama releases omit the gfx906 support these cards need, so the stack is llama.cpp on ROCm. Each card is power-limited to about 150 W, because generation speed is bound by memory bandwidth, not compute.

The parts come to about €3,530, with the GPUs bought used at about €220 each. Build notes: Introducing RAIV.

How to use this page

These are bills of materials. Source the parts and build the box, or hand the list to an integrator who will. The wirken binary does not know the difference between a reference build and a comparable machine you assembled yourself; it asks the OS for a vault key, a Unix domain socket, and a network interface, and Spule, Fern, and Heim all provide those the same way as any other Linux or macOS host. RAIV runs no wirken at all; wirken reaches it as an inference endpoint. If you have a stricter security target than these references aim at, the components that matter to a wirken deployment are the same components a hardened workstation already gets right: full-disk encryption, a real TPM, secure boot, and an OS keychain you trust to hold the vault key.