SqueezeBits, a Seoul inference optimization shop, delivered quantized open-weight vision models to POSCO DX so steel mills can flag surface defects on line-side GPUs without uploading slab images to public clouds, the partners said Monday.
The deployment targets hot rolling lines where latency budgets are measured in milliseconds and privacy rules treat production imagery as trade secrets. POSCO DX, the IT affiliate of the steel group, tested SqueezeBits’ toolchain on open convolutional and transformer backbones, then distilled them into INT8 pipelines that fit NVIDIA edge cards installed near pickling tanks.
Why quantization mattered on the mill floor
Full-precision models trained in POSCO’s central data lake were accurate but too slow for real-time alerts. Sending frames to a remote API introduced jitter when plant networks contended with operational technology traffic. Quantization reduced memory footprint enough to run two models in parallel—one tuned for scale defects, another for edge cracks—while keeping inference under 40 milliseconds per frame in pilot logs reviewed by InfoHandle Network.
Open weights versus bespoke training
POSCO DX chose open weights to avoid vendor lock-in and to simplify audits. SqueezeBits applied structured pruning and calibration datasets drawn from anonymized historical defects, but the base architecture remains documented in open repositories. Plant engineers can retrain heads locally when product mixes change, a requirement union safety committees insisted on before cameras expanded beyond the pilot bay.
The Ministry of Science and ICT has promoted on-prem AI for critical infrastructure; steel qualifies under guidance issued this summer. Subsidies did not fund this project directly, but POSCO DX cited the policy when explaining why cloud-only vendors were dropped from the shortlist.
Human oversight remains
Vision alerts still route to veteran inspectors on hardened tablets. The models do not auto-reject coils; they color-code anomalous regions for human confirmation. POSCO DX said false positives dropped versus rule-based vision from the prior decade, but teams maintain override buttons when maintenance lighting skews cameras.
Expansion plans
If the Pohang pilot holds through winter humidity swings, POSCO DX will replicate racks at Gwangyang and consider exporting the stack to joint ventures overseas. SqueezeBits hopes the case study attracts other heavy industry clients—shipbuilding plate lines and battery foil plants—with similar edge constraints.
Competitors including cloud hyperscalers and Korean systems integrators offer managed vision, but POSCO’s choice signals preference for on-site control when molten metal is in frame. The next test is durability: models must run months without centralized retraining, proving that compressed open weights can live in factory heat and dust without drifting.
Maintenance and model drift
Edge GPUs in mill environments face dust and vibration. POSCO DX installed filtered enclosures and vibration-damped mounts after a prior camera failure speckled training data. SqueezeBits ships checksum tools so plant IT can verify quantized weights were not tampered with between updates. Monthly drift checks compare live defect counts against human audit samples; if precision falls below agreed thresholds, models roll back to prior versions automatically.
Union safety reps won the right to pause automated alerts when maintenance crews work within camera sightlines, preventing false stops during legitimate equipment servicing.
POSCO DX executives told plant managers that savings from fewer misgraded coils should fund additional edge racks if payback stays under three years. Competing integrators may bid on phase two, keeping SqueezeBits honest on pricing for quantization services.
Export customers auditing POSCO mill certificates may ask whether vision-assisted inspection changes defect classifications; POSCO DX said human sign-off remains on all release documents for now.
