AI chips are designed to process AI tasks, however their usefulness has been limited by their high power consumption. By tackling this problem through algorithm and architectural optimization, Professor Zhou Jun and his colleagues at the University of Electronic Science and Technology of China (UESTC) decreased power utilization.
Two artificial intelligence (AI) chips with record-breaking performance have been unveiled by scientists. These artificial intelligence (AI) microchips are said to be the most energy-efficient in the world by scientists. They can be installed inside of smart devices because they are so tiny.
To allow voice control, the first chip is made to fit inside of smart devices. Because it can identify the audio signals of a target speaker, this device is excellent for speaker verification and keyword identification.
This new AI chip’s primary benefit is its capacity to get around the drawbacks of conventional voice recognition software. This chip, in contrast to conventional systems, can detect speech from the target speaker with accuracy even in noisy settings when there are distractions like music, TV, or conversations. This feature improves its dependability and efficacy in practical situations.
With less than two microjoules needed for each recognition, the device demonstrates remarkable energy efficiency. With an accuracy rate of over 95% in quiet situations and 90% in loud conditions, it creates new global standards for energy efficiency and voice recognition technology.
The tiny one sq cm (0.155 square inch) chip was effectively incorporated into a 3 cm x 3 cm microcontroller unit inside a toy car during a system demonstration, providing exact control over its movements.
Among the low-power voice control applications that the chip might be used for include smart homes, wearable technology, and intelligent toys. Its compact design and low power consumption make it ideal for delivering dependable voice control capabilities to a wide range of devices.
The second gadget the researchers brought to the conference is meant to detect seizure signs in epileptics. The device, which was specifically created for wearable technology, employs electroencephalogram (EEG) detection to detect ongoing epileptic seizures, enabling patients to receive therapy or emergency care as soon as possible.
Owing to the rarity of seizures and the need for hospitalization, a significant amount of patient seizure data has historically been needed for training in order to achieve high accuracy in seizure detection. This method is costly and time-consuming. To get around this problem, the researchers developed a zero-shot retraining technique. By refining a pre-trained AI model to reliably forecast seizures in unseen data—all without obtaining seizure signals from patients—this method attains an astounding accuracy rate of over 98%.
Patients simply need to calibrate the device for two minutes before utilizing it in a natural way. The gadget can now recognize different signal characteristics thanks to this calibration, which enhances its capacity to recognize epileptic episodes.
Thanks to advances in feature extraction and on-chip learning algorithms, this device achieves an impressive average recognition energy consumption of about 0.07 microjoules, making it the most energy-efficient design globally.
Comparing the official report to a chip that was showcased at the previous year’s conference, it shows an impressive 90% reduction in energy consumption and a notable 10% gain in accuracy.
In addition, the device may find application in fields other than seizure detection, like brain-computer interfaces and sleep monitoring.
During an ISSCC demonstration, real-time EEG readings from a wearable brain-computer interface device were relayed over Bluetooth to the test board. The tool was altered to recognize fictitious motor commands, enabling the operation of a robot’s forward, backward, and stop motions.
At the 2024 IEEE International Solid-State Circuits Conference (ISSCC), the integrated circuit (IC) industry’s Olympics, the team showcased two of these cutting-edge circuits.
For More Visit: https://munafekideal.com/