AIBIZ, a South Korean industrial AI startup, runs its DutchBoy software on AMD EPYC 9355 and 9554 server CPUs, without GPU acceleration, to catch semiconductor wafer defects before they cause failures. According to AMD's case study, AIBIZ saw a 30% AI inference performance increase over its previous Intel CPUs on five domain datasets, and CTO Hyun Jin Choi says DutchBoy allows a yield improvement of 3% to 5%.
“We can find small issues at the beginning of the fabrication process to increase the yield of the foundries,” says Hyun Jin Choi, CTO, AIBIZ. He explains that quality checks are usually done at the very end of production, so defects are detected at a late stage, while sensor data lets DutchBoy identify problems as they occur.
The server AIBIZ works with is installed within the Samsung Electronics plant inside the etching equipment itself. It collects real-time sensor readings from 20 pieces of equipment, at a sampling rate of 100 milliseconds from 300 sensors in each. DutchBoy runs in a Docker container-based system, with 20 AIBIZ products connected to one server, and uses a Time Series Anomaly Detection AI analysis model and a graph neural network (GNN) that analyzes correlations between sensors such as the temperature and pressure sensors inside the etching chambers.
That consolidation lets AIBIZ keep its models below 100,000 parameters, compared with up to 120 billion for a typical current LLM. Choi says very large parameter counts require GPUs, which generate more heat and need extra cooling inside a fab, and that GPU-to-CPU latency is also an issue. GPUs are still needed for training, but inference is within the capability of a fast CPU: from collection through pre-treatment, inference, visualization and reporting, five minutes is fast enough.
AIBIZ moved to AMD after bottleneck problems with its previous Intel processors; it was looking for strong multithreading. AMD provided initial test servers, and hardware partner HPE helped identify which EPYC product would be best. When arcing occurs inside a chamber, the data spikes, and DutchBoy provides real-time information to engineers so they can prevent the wafer defect.
Choi estimates that memory wafers cost about 20 million won ($13,500) each and that, before AIBIZ's technology, companies sampled one in every 20 wafers; finding a problem before it happens can save 20 wafers, or 380 million won ($255,000), which he multiplies over 365 days to 150 billion won ($100 million) per year. AIBIZ currently uses the processors at Samsung and Xian (China) plants and plans to use the same processors for outsourced semiconductor assembly and test (OSAT) facilities and LG Innotek, and to expand to Hynix, Intel and Micron using Samsung as a reference.
Details