Chinese scientists have unveiled two ultra-low-power artificial intelligence (AI) chips with record-breaking performance at the most prestigious conference in the chip design industry. AI chips, specifically engineered to process AI tasks, typically require substantial power due to the heavy computational demands they face, which has limited their application in real-world scenarios. But through algorithm and architectural optimization, Professor Zhou Jun and his team from the University of Electronic Science and Technology of China (UESTC) have managed to significantly reduce this power consumption. The team presented two of these innovative chips at the prestigious IEEE International Solid-State Circuits Conference (ISSCC) 2024. The ISSCC is an annual global gathering for solid-state circuits, where the best researchers, engineers and professionals meet to talk about new developments and the future of chip technology. This year’s conference was held in San Francisco from February 18 to 22.
The first of the two AI chips was designed to be embedded into smart devices to enable offline voice control. This chip excels in keyword spotting and speaker verification by recognizing the voice signals of a target speaker, even with environmental noise, such as television, music or other people talking, which typically interferes with traditional voice recognition chips. Conventional chips also tend to suffer from high wake-up energy consumption and frequent false wake-ups, leading to low energy efficiency. Zhou’s team proposed a novel architecture that overcomes these limitations through multiple optimizations, including dynamic computation engines, an adaptive noise suppression circuit, and an integrated keyword and speaker recognition circuit. The chip also has applications in low-power voice control scenarios such as smart homes, wearable devices and smart toys.
The second chip the team presented at the conference was designed to detect seizure signals for people with epilepsy. Created to be used in wearable devices, it uses electroencephalogram (EEG) recognition to alert that an epileptic seizure is ongoing so the patient can seek medical help or treatment. Existing designs rely on extensive patient seizure data for training to achieve high accuracy, a process that is time-consuming and costly due to the low occurrence of seizures and the need for hospitalization. To solve this, the researchers optimized a zero-shot retraining algorithm allowing a pre-trained AI model to make accurate predictions on unseen data without the need for collecting patients’ seizure signals, achieving an accuracy rate of over 98%. This chip also has potential applications beyond seizure detection, including other brain-computer interfaces and sleep monitoring, the South China Morning Post reports.
