1. The protocol under the pavement

Neal Stephenson’s Snow Crash moves between the physical and virtual worlds, but the real action often depends on who controls the underlying protocols. The lesson for the desktop is similar: the data path across the attachment interface is small in physical size but large in consequence. If the cells cannot talk, the federation does not exist.

This entry is about how data moves across the cell-to-rack interface.

2. What the data path must carry

A cell may need to communicate:

  • Management traffic: temperature, voltage, current, health status.
  • Compute traffic: storage access, inter-process communication, customer payloads.
  • Control traffic: commands to power cycle, reset, or reconfigure.
  • Timing: synchronization between cells for coordinated operation.

These have different requirements. Management traffic is low bandwidth and high reliability. Compute traffic may be high bandwidth. Timing traffic is low bandwidth but low latency and low jitter.

3. Physical layer options

The data path could use:

  • Ethernet over copper: high bandwidth, well understood, but sensitive to EMI and connector quality.
  • CAN bus or similar: robust for management, lower bandwidth.
  • PCIe: very high bandwidth, but more complex and less tolerant of hot-swap.
  • Optical: immune to EMI, but adds complexity and alignment requirements.

For a first-generation desktop, Ethernet for compute and a robust management bus is a reasonable split. The attachment connector would carry both.

4. The hot-swap problem

The hard part is not the wire. It is the protocol behavior when a cell appears or disappears. The network must:

  • Detect insertion and removal.
  • Assign or reassign addresses.
  • Re-route traffic around a removed or failed cell.
  • Maintain consistency for storage and state.

This is where the weak federation model is tested.

5. What the first test should measure

The ground experiment should at least verify that:

  • The connector’s data pins mate reliably.
  • The link comes up after insertion and goes down cleanly after extraction.
  • Bit-error rate is acceptable.
  • Hot-swap does not crash the remaining network.

Full protocol testing can come later; the first test only needs to prove the physical layer is sound.

6. What this changes

  • The attachment interface is a data network interface as well as mechanical, thermal, and electrical.
  • The data path design should separate management and compute traffic.
  • The first ground experiment must include basic link-up and link-down behavior.