The Lab runs its own metal. The Spark Duet, a two-node NVIDIA DGX Spark cluster, carries the serving, training, and creative load daily; the workstation and networking discipline around it is the buildout capability we bring to clients. We publish topology and measured behavior, never endpoints or credentials.
TWO NODES · ONE LABThe Lab runs its own metal: the Spark Duet, a two-node NVIDIA DGX Spark cluster, plus the workstation and networking discipline around it. We publish topology and measured behavior, never endpoints or credentials.
The register lists what is live and what is on the roadmap, including the board-design program that is open for sponsorship.
Two NVIDIA DGX Spark units operated as one lab: a pair of GB10 Grace Blackwell machines with 128 GB of unified memory each, joined over a 200 Gb/s ConnectX-7 link. One node serves the gateway, models, and agents; the other trains, renders, and monitors. The pair is operated daily, measured honestly, and documented in public.
Two NVIDIA DGX Spark units operated as one system: each a GB10 Grace Blackwell machine with 128 GB of unified memory, joined over a 200 Gb/s ConnectX-7 link. Small enough to sit on a desk, large enough to serve, train, and render for a working lab.
Node 1 serves: the OpenAI-compatible gateway, the model lanes, the agents. Node 2 trains and renders: QLoRA fine-tunes, ComfyUI pipelines, batch jobs, and the monitoring that watches both. Serving stays stable while the second node saturates.
A pair buys role separation, a 256 GB combined ceiling for models that need tensor parallelism across the fabric, and a failure story: either node can carry the essential services alone while the other is down for maintenance.
The pair is operated through SparkDuet, the Lab's open source operating layer: one endpoint, four serving lanes, topology switches in one command. Its register entry is LA/01 and its page carries the measured numbers.
Custom AI workstations, GPU systems, storage, and networking for local AI labs, designed and built as an engagement. Scope runs from a single fine-tuning workstation to a small serving cluster with its monitoring stack.
Specify, benchmark, operate, publish. We recommend only what we have run: the decision records behind the Lab's own purchases, with their measured behavior, are the template a client engagement starts from.
The research direction beyond buildout is board-level design for local AI sensing: vision and edge inference carriers that pair with the local stack. The Lab is open to a sponsor or design partner for that program; the register will carry it with an honest status when work begins.
If your team needs a local AI lab specified and built, or wants to sponsor the sensing program, write to [email protected] or use the briefing form. Roadmap means roadmap: we will say plainly what exists today and what does not.
The Spark Duet: two NVIDIA DGX Spark units, each a GB10 Grace Blackwell machine with 128 GB of unified memory, joined over a 200 Gb/s ConnectX-7 link. One node serves the gateway, models, and agents; the other trains, renders, and monitors. The open source operating layer for the pair is SparkDuet.
Two nodes buy three things one box cannot: role separation (serving stays stable while training saturates the other node), a 256 GB combined memory ceiling for models that need tensor parallelism across the ConnectX-7 fabric, and a failure story where one node can carry the essential services alone.
Workstation and lab buildout is on the register as a roadmap item, and design engagements are open. The method is the differentiator: we specify, benchmark, and operate the hardware we recommend, and the decision records behind our own purchases are the template an engagement starts from.
Topology, node roles, counts, and measured behavior are public. Addresses, ports, hostnames, credentials, and anything else an attacker could use never reach a page, a repository, or the site concierge.
Board-level design for local AI sensing (vision and edge inference carriers) is on the roadmap as a research direction, and the Lab is open to a sponsor or design partner for that program. The register will carry it with an honest status when work begins.
We don't pitch slide decks. We show you what we've already built in your domain, then engineer what your mission requires.
