Answers are grounded in this paper and cite their sources.
The identification and testing of semiconductor components is among the most expensive stages of memory production. Despite multi-insertion testing at several temperatures, customers still report application failures of 20–50 ppm for complex components such as dynamic random-access memory (DRAM), and field failures are often not reproducible in the production flow (no trouble found, NTF). This work examines whether application-based testing under application-representative conditions can expose such failures. Two platforms are used: an FPGA-based tester (RHINOE) with 192 sockets on six tiles, and an Intel CPU-based platform (OCTOPUS) that reproduces a customer memory-controller and BIOS environment. For about 10,000 4 G DDR3 devices, the conventional array test gave about 2% yield loss, whereas the FPGA-based application test gave about 6% at the same 98 °C/110 ms retention guard band. Five of ten customer-returned DDR3 devices, whose failures were not reproduced in the conventional flow, failed on the FPGA platform. In a 32-device cross-tile comparison, five devices that passed under calibrated settings failed when per-tile voltage calibration was omitted. In a one-week test of 1000 DDR4 devices, two transient single-cell events were observed. The results support application-based testing as a complement to production test; a zero-defect rate is not demonstrated, and the mechanisms of the transient events were not identified.
No takes yet. Share an insight, caveat, or question.
Yin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: