FOG-NODE-PT-CM-001 · proposed R&D · no public date

Post-MacBook Fog appliance

From the live MacBook laboratory host to a standalone, always-on mini-server. Modular compute, resource evidence, and energy-aware hardware. Not a production device.

Canonical text: docs/FOG-APPLIANCE.md. Software today remains v0.6.0-lab on the Mac. Experimental thermal cartridges must not be single points of failure.

Master roadmap Infrastructure hut Full prospectus Fog health

01 · proposition

The Fog is planned capacity

Edge contributes residual capacity. Fog is installed, persistent, and provisioned on purpose. A useful node is not a peak-benchmark machine. It is sustained, measurable, verifiable work per euro, watt, cubic centimetre, and maintenance hour.

Nominal watts are not operating cost. Operating cost is compute + memory + storage + network + conversion loss + cooling + control + standby. Declared capacity is not accepted capacity. The hardware must expose that difference.

02 · first prototype

Gen-0 is a used mini-PC

Do not manufacture a custom appliance first. Target about €50–100 second-hand: x86 mini-PC or thin client, ≥8 GB RAM, working SSD, Ethernet, Linux, conventional cooling, a cheap MCU, basic temperature and power sensors.

Success is not a new cooler. Success is: another machine sends work, gets a useful result, and can verify that this node provided the claimed resource for the claimed interval.

Used mini-PC ├── Linux host ├── StrataMesh Fog runtime ├── resource collector + benchmark ├── evidence / receipt engine └── hardware telemetry daemon └── USB/serial MCU ├── temperature ├── power └── watchdog

03 · replaceable cartridges

Nine modules

A Compute

Board, CPU, RAM, optional NPU/GPU/FPGA, local fabric.

B Memory / storage

RAM, NVMe, node state, evidence store, future CXL.

C Network

Ethernet first. Later programmable NIC / DPU / FPGA.

D Evidence / security

TPM or secure element, MCU, RTC, sensors, watchdog.

E Conventional thermal

Cold plate, heat pipe or loop, ordinary fan. Must work alone.

F Experimental thermal

Magnetocaloric cells, regenerator — research only.

G MEMS air

Piezoelectric movers for hot-side rejection.

H Energy electronics

DC bus, rectifier, supercap, measured recovery only.

I Accelerator

FPGA / NPU / CIM / photonic — optional cartridge.

The node stays up if F, G, H or I are removed.

04 · evidence plane

Declared vs measured

Software may say “32 cores.” Hardware-backed evidence says this identified device booted measured firmware and completed the tests. The OS must not freely rewrite MCU measurements.

SERVICE PLANE workloads RESOURCE PLANE CPU RAM disk net accel EVIDENCE PLANE measure → timestamp → hash → sign SECURE MCU signed receipt

Time is a measurement primitive: start, end, allocation, work done, sensor health, clock drift. Graduated capacity from 100% down to safe shutdown — thermal headroom cuts advertised availability.

05 · thermal / energy

Conventional first

Copper cold plate and a normal heatsink/fan are the primary path for several generations. Magnetocaloric work starts as a separate 20–50 W bench, not 300 W on the node. MEMS air rejects heat to ambient; it is not below-ambient refrigeration. Thermomagnetic recovery is counted only as measured watts at the electrical interface. The node is not self-powered.

Net demand = electrical input − measured recovered output. Dew-point interlock before any below-ambient experiment.

06 · ladder

Phases A–J

A · cheap reference

Used mini-PC + Linux + current Fog stack. Execute remote work. Report contribution.

B · instrumented

Sensors, MCU, watchdog, benchmark, tamper-evident log. Nominal ≠ sustained.

C · dedicated appliance

Replace the temporary box. Equal or better useful capacity than the MacBook at lower TCO.

D · data plane

FPGA/DPU/NIC metering when justified by measurement.

E · accelerator cartridge

One real StrataMesh workload must benefit.

F–J · MEMS / MCM / recovery / integrate / density

Only after measured advantage. Long runs 30 / 90 / 180 days.

07 · what wins

Useful work per euro and per kWh

Over the useful life of the node — not colder CPUs, not theoretical recovery, not TOPS. Do not start with custom silicon, superconducting magnets, a large GPU before workloads, or an enclosure before interfaces exist.

08 · Gen-0 spec

Compact Fog — planning range

Compute: efficient used x86 (ARM allowed; x86 first for current tooling)
Memory: 8–32 GB Gen-0; 32–128 GB later
Storage: one SSD/NVMe; separate evidence store when practical
Network: integrated Ethernet; 2.5 GbE if cheap
OS: Linux host for the same Fog role the Mac runs today
Thermal: conventional + instrumentation
Identity: software first; TPM-class next revision

Path: MacBook reference → minimal Fog → instrumented → evidence → programmable → solid-state thermal → energy-aware → composable appliance.