The First Machine: What One Spark Actually Does

Chapter 02

A single object on private land

Hearthmind starts with a single object: one NVIDIA DGX Spark sitting on a piece of private land.

On paper, it looks like marketing hyperbole: a personal AI supercomputer promising one petaFLOP of FP4 performance and 128 GB of unified memory in a 240 watt box. In practice, it becomes a way to answer a simpler question: what can one Spark actually do when you treat it as part of the property, not as a fancy office PC?

Chapter 02

The hardware baseline

GB10 Grace Blackwell combines 20 Arm cores — ten high-performance Cortex-X925 and ten efficient Cortex-A725 — with a Blackwell GPU and 128 GB of LPDDR5x on a 256-bit bus at up to 273 GB/s. Storage is a 4 TB NVMe SSD. Networking is 10 GbE to the LAN and dual QSFP via ConnectX-7 for up to 200 Gbps links to future nodes.

Chapter 02

The experiments

From that, Hearthmind defines a set of experiments:

• Local inference: how fast can it serve modern 8B, 70B and 140B language models under realistic prompts?
• Private RAG: can it index and search the Solystopia archive locally, answering questions from documents that never leave the property?
• Agent workflows: how well can it coordinate research agents, coding tools and household automation?
• Fine-tuning: where is the line between adapter-based fine-tuning being feasible and full training needing a cluster?

Chapter 02

Why storage appears immediately

On day one, Spark lives on local SSD. Almost immediately, that proves inadequate. Model weights, embeddings, archives and backups compete for a single 4 TB volume. So the first side quest appears: Hearthvault 01, a separate storage server with ECC, ZFS and 10 GbE designed to hold models, datasets and archives sustainably.

Chapter 02

The machine as architecture

What one Spark actually does, then, is less about AI benchmarks and more about architecture. It shows that datacentre-class inference and fine-tuning can genuinely live on a property; it makes storage, networking, energy and thermal systems part of the intelligence design; and it provides a baseline against which every subsequent stage can be measured.

Hearthmind’s content for this stage is not a synthetic review. It is a logbook: exact specs and UK price paid, measured throughput for specific models and tasks, energy usage per workload type, heat behaviour in different seasons, and the ways in which the machine falls short.

Chapter 02

Everything grows from here

Only once that evidence exists does the project move on to the next questions: What is the smallest workshop-scale cluster that can supply most of a household’s useful intelligence? How does it behave when tied to PV and batteries? How much of its waste heat can be reused?

The first machine is how Hearthmind learns to treat AI as infrastructure. Everything else grows from there.