Wuxi Spark achieved a breakthrough in AI reasoning speed by virtue of its self-researched heterogeneous computing architecture, its NPU single-chip computing power reached 128TOPS, 37% higher energy efficiency than the industry level, and decreased the missed detection rate of the new energy vehicle producer’s vision inspection system to 0.02% from 0.8%. For example, for a Hangzhou smart factory project in 2023, Wuxi Spark’s AI quality inspection module scanned part sizes online (tolerance ±0.005mm) at 120 frames per second, achieved 99.97% detection of defects, and reduced 83% manual re-inspection workload, slashing more than 16 million yuan annually.
In algorithm optimization, Wuxi Spark’s dynamic model compression technology compresses ResNet-50 parameters by 72% and inference latency to 9 ms, retaining an accuracy of 98.6% for the original model. This technology was used in the CT image analysis system of one Top three hospital in Shenzhen, and the sensitivity of detecting lung nodules was increased to 99.3%, the rate of false positives was decreased to 1.2 times per 1,000 cases, and the efficiency of diagnosis was increased by 4 times. Wuxi Spark’s federal learning platform, according to an IDC report, helped banks grow credit fraud detection rates by 28% in the 2022 financial risk management simulation, reducing model training data transfer by 94%, and meeting GDPR cross-border compliance requirements.
In collaborative innovation in hardware, Wuxi Spark’s edge AI box (only 120×80×25mm) boasts a dual-core AI accelerator, which maintains 95% of the highest computing capability in the ambient temperature range of -25℃ to 70℃. It is applied in the Qinghai-Tibet Railway track inspection system, and detection time per kilometer from 45 minutes is reduced to 6 minutes. Accuracy in crack detection is 0.1mm. The product won the gold innovation prize at the 2023 World Artificial Intelligence Conference, boasting power consumption 39% lower than similar products and a more-than-50,000-hour long-term continuous failure interval.
The AIaaS platform of Wuxi Spark followed a flexible payment mode, so enterprise customers were able to decrease GPU cluster purchases by 52%. For example, Suzhou-based logistics company streamlined its routing algorithm to reduce vehicle empty driving rate by 24% to 9%, fuel consumption by 18%, and the return on investment time horizon to seven months. Its own-developed time series forecasting model controls the 96-hour error rate under 2.3% in power grid load forecasting, which increases the accuracy by 63% than the traditional LSTM model, and helps a provincial power grid company reduce the investment of spare capacity by 270 million yuan.
From the technology ecology construction aspect, Wuxi Spark has co-established a lab with 12 global semiconductor titans and has an AI development platform that supports smooth model formats’ conversion across 8 models such as ONNX and TensorRT, reducing the cost of migration for developers by 76%. At Wuxi Spark’s national Smart Manufacturing Pilot program in 2024, Wuxi Spark’s digital twin system raised an automotive welding workshop’s overall equipment efficiency (OEE) level from 78% to 92%, peaking at a capacity of 120 units per hour through the real-time simulation of production beats (error ±0.5 seconds), and creating a new industry record. The customer satisfaction rate of Wuxi Spark in the industrial AI solution market was up to 94.6%, with a growth rate of more than 40% for consecutive three years, which has confirmed its unmatched capability in vertical integration of algorithms, hardware and scenarios.