GENEVA Investments in time and dollars notwithstanding, semiconductor soft errors remain a tough nut to crack.
Certainly, scientists and engineers have known for decades that soft errors, which occur as a result of cosmic radiation, can alter the state of the circuit, causing systems to malfunction unpredictably. The industry has even jury-rigged a few remedies to shrink soft error rates (SERs), applying them in DRAMs, SRAMs and chips targeted at space-borne applications.
But the accuracy of current methods for SER modeling and test is a lingering unknown. Most manufacturers jealously guard test results, making it tough to draw useful comparisons and conclusions on SER data collected so far. And nobody knows how well current SER predictions will hold up under the next process node's finer geometries.
Some clues could be forthcoming this week, when an emerging test program announces its preliminary results in Athens at the Radiation Effects on Components and Systems (Radecs) technical conference. Since March, STMicroelectronics and the Laboratory for Materials and Microelectronics of Provence (L2MP) have been quietly conducting real-time SER testing of semiconductor memories at a site in the French Alps, at an elevation of 2,552 meters, under Europe's Altitude Single-Event-Effects Test European Platform (Astep) program. Intending eventually to open its facility to the industry, Astep has also brought in Xilinx FPGAs to test.
Through real-time soft-error testing at high altitudes, Astep hopes to add another data point to the established suite of approaches for estimating SER. Those current methods include accelerated alpha-particle and neutron or proton source/ beam testing; life testing in natural environments; and modeling and software simulation at the device or circuit level.
An important advantage of high-altitude facilities is that "upsets happen from five to 15 times more often, depending on altitude, thus generating better data, sooner, with less uncertainty," said Austin Lesea, principal engineer at Xilinx Inc.
"Use of high-altitude testing stations more closely mirrors a 'real world' application, compared with bombarding a sample with radioactivity to simulate the conditions a part would see in the real world," said Richard Wawrzyniak, senior analyst at Semico Research Corp.
Additional real-time SER testing is "critical to giving our customers scientific confidence in the robustness of our chips," said Jean-Pierre Schoellkopf, director of advanced design, central CAD and design solutions at STMicroelectronics. Astep's mission, he said, is to prepare for what SERs may look like "10 to 15 years from now."
Leaving no trace
SERs are "a tricky phenomenon to study and model," said Olivier Lauzeral, general manager at iRoC Technologies Corp. Occurrences are rare, making SERs difficult to observe. Worse, "they don't leave any physical trace in the chip that could allow for postmortem investigations," Lauzeral said.
SER testing deals with many probable events, including the "locations of the collision of particles with silicon atoms, the energy involved, the by-products created and the charges deposited," Lauzeral said. The litany of potential unknown factors tends to render SER predictions and even artificially accelerated tests "prone to errors," he said. Maximizing prediction accuracy thus requires "a large amount of test data points, correlated with simulation and a comparison with real-time, nonaccelerated testing," Lauzeral said.
Bob Patti, chief technology officer of Tezzaron Semiconductor Corp. (Naperville, Ill.), lamented the lack of modeling tools to gauge SER interactions in materials. Further, he said, current testing doesn't test circuits at speed. Citing the generally slow, FPGA-driven speeds for SER testing, he said, "If the circuit is given enough time to recover, the error just disappears. We can't model, and the testing methodology is flawed."
Currently, the most widely applied approach to characterizing SER uses alpha-particle sources and neutron beams with high intensity. In a recent paper, Xilinx researchers wrote that "beam testing is not accurately predicting the real improvements that are being realized in atmospheric testing."
"Actual atmospheric data is absolutely essential," Xilinx's Lesea said, "as beam testing has now been shown to be pessimistic by factors as great as five to eight."