1. The logic that was not supposed to be smart

Murray Leinster’s “A Logic Named Joe” returns as the recalled work because the story’s accidental super-logic is still, physically, an ordinary logic. Its special behavior comes from a configuration mistake, not from a different kind of hardware. The desktop’s compute problem has the same shape: the physical computers are mostly ordinary, and the reliability must come from how they are arranged, watched, and recovered.

This entry reads about COTS compute and radiation tolerance in LEO.

2. The flight-heritage case for ordinary computers

The most direct evidence that ordinary computers can work in orbit is HPE’s Spaceborne Computer line:

  • Spaceborne Computer-1 flew to the ISS in 2017 and returned in 2019 after roughly 657 days of continuous operation. The hardware was essentially off-the-shelf HPE Apollo servers, hardened in software rather than by custom radiation-hard parts. One of four power supplies failed and nine of twenty SSDs developed issues, yet the system kept running through redundancy and recovery (The Register).
  • Spaceborne Computer-2 launched in 2021 with GPUs and an Azure edge link, running AI, genomics, and image-processing workloads. It returned in January 2023 and was followed by an updated unit. The central claim remained the same: software monitoring, thermal throttling, error detection, and autonomous recovery can substitute for rad-hard hardware on short-to-medium LEO missions (HPE).

The lesson is not that COTS hardware is immune to space, but that its failures can be made survivable if the system is designed to detect them and keep working around them.

3. The radiation environment in numbers

LEO is not a clean room for electronics. A 2024 Tsinghua-led survey of COTS in-orbit computing notes that space radiation causes single-event upsets (SEUs) and multi-cell upsets (MCUs) at rates that matter for ordinary memory. For a typical 40 MB memory footprint, the paper estimates roughly 150 bit errors per day, with more than 5 % of those appearing as spatially correlated multi-bit errors (Wang et al.).

The South Atlantic Anomaly is the worst region, but the problem is global and continuous. The bit errors are often silent: a DNN inference degrades rather than crashes, a telemetry value drifts rather than raises an alarm. This makes detection harder than with a clean hardware fault.

4. Three protection strategies

The literature groups protections into three layers:

  • Hardware protection: radiation-hardened processors, triple modular redundancy, and ECC memory. This is the traditional approach. It works, but it costs more, lags terrestrial performance by years, and consumes more mass and power.
  • System software protection: watchdogs, reboots, page-level checksums, and software ECC. These are lighter but can multiply runtime or memory usage and may not catch silent errors in applications.
  • Application-aware protection: the Wang et al. argument. Not all bits are equally important to a given workload. In DNN inference, for example, errors in shallow layers propagate and amplify, while errors in some deeper layers are damped by redundancy. Designing the model to suppress error propagation and allow early exit can reduce the impact of radiation errors to near zero with modest overhead.

The desktop is not a DNN satellite, but the principle generalizes: understand what the workload actually needs, then protect that rather than every bit equally.

5. What this changes

  • COTS compute in LEO is viable if failures are detected and contained, not prevented outright.
  • Radiation-induced bit errors are frequent enough that any long-duration desktop must plan for them.
  • Protection can be layered at hardware, system software, and application levels; the cheapest layer depends on the workload.
  • The next entry will read about storage, which is where the HPE experiments saw the most casualties.