In the 1970s, Intel had a failure it could not explain. DRAM chips were corrupting data in the field, and the corruption followed no pattern. A memory cell that should have held a 1 would read as a 0. The chip would pass every test in the factory, work fine for hours, and then quietly hand back the wrong answer. Engineers chased the problem through manufacturing, through the chip design, through the rest of the system, and nothing matched the symptoms.

The cause was sitting next to the silicon the whole time. The ceramic package that encased the DRAM was emitting alpha particles. The particles were striking the die and flipping bits. The material meant to protect the chip was the thing destroying its data.

The mechanism of a bit flip

The 1-to-0 flip is the signature of the event. DRAM stores each bit as a tiny electrical charge in a microscopic capacitor. A charged cell is a 1. An empty cell is a 0. The arrangement is cheap and dense, which is why DRAM became the standard memory technology of the era. The trade-off is fragility. The charge in a cell is small, and it leaks away even in normal operation, which is why DRAM is called dynamic and why the chip must constantly refresh every cell.

That smallness is also what makes a cell vulnerable to a single stray particle. An alpha particle is a helium nucleus: two protons and two neutrons bound together. It is emitted by radioactive elements as they decay, and the elements that emit it, uranium and thorium among them, are found in trace amounts in many natural materials. Clay is one of those materials, and clay is the base of most ceramics, including the kind used to package chips. A package can contain specks of these elements too small to see and too dispersed to matter in any ordinary sense. The package is not hazardous and does nothing visible. It contains a few atoms that will, sooner or later, decay.

The decay of one of those atoms sends an alpha particle out of the nucleus at high speed. If the particle reaches the silicon die, it tears through the crystal lattice and strips electrons from atoms along its path. In the neighborhood of a memory cell, those freed electrons change the charge balance. A cell holding a charge can lose enough of it to read as empty. That is a 1 flipping to a 0. A cell that was empty can also collect the stray electrons and read as full, which is a 0 flipping to a 1. The stored data is wrong either way.

A failure with no schedule

The reason the problem stayed hidden is that the event is probabilistic. Radioactive decay does not run on a schedule. One package might emit a particle in a day. Another might sit quiet for months. The failures were scattered across chips and systems with no rhyme or reason, and they looked like every other kind of intermittent hardware trouble. A chip that failed in the field would often pass every test in the lab, because the tests ran for minutes while the particle arrived on an entirely different timescale.

This is what engineers now call a soft error. The term separates two very different kinds of failure. A hard error is a broken component: a burned-out trace, a cracked die, a shorted junction. It does not recover. A soft error is a transient event. The hardware is undamaged. The chip operates normally the instant after the bit flips. But the data is already corrupted, and the system has no way to know which bit is wrong.

Soft errors were not a familiar category in the 1970s. The working assumption was that a memory chip that failed must be physically defective, and the idea that a fully functional chip could return a wrong answer because of a particle emitted by its own casing was not part of the playbook. That blind spot is why the investigation dragged. The packaging looked inert. It was hard, stable, and insulating. It was not radioactive in any way a person would notice. But an imperceptible level of radioactivity is not the same as no radioactivity, and the difference was enough to corrupt data.

The problem never disappeared

The case made clear that a chip's package is not a neutral shell. It is part of the electrical environment of the device, and its material properties, including its radioactive purity, matter as much as its shape and its ability to conduct heat. The discovery also gave the industry a vocabulary and a mechanism for a class of failure that never went away. Alpha particles still come from packaging materials that are not pure enough, and other radiation sources have joined the list. High-energy cosmic rays striking the atmosphere produce secondary particles that reach ground level and flip bits in the same way. Every modern computer system sits in a weak but constant hail of these particles.

Hardware designers have spent the decades since managing the odds. Error-correcting code memory, common in servers and workstations, detects and fixes single-bit flips before they reach the software. Chip makers screen materials for radioactive contamination. None of this eliminates the problem. It shrinks the odds, and that is the only realistic goal when the enemy is a random decay event inside the material you built the chip out of.

The Intel case from the 1970s remains the cleanest example of why this matters. The failure looked like a manufacturing defect and behaved at times like a software bug. The mechanism was direct: a helium nucleus, fired from inside the ceramic shell, arrived at the worst possible moment and turned a stored 1 into a stored 0.