According to SCMP, a research team from Tsinghua University in Beijing (China) published a report, introducing the Tianmouc chip with a record image processing speed. They said that the chip can collect image information at a speed of up to 10,000 frames per second, with 10-bit accuracy, and a dynamic range of 130 decibels. In addition, Tianmouc also reduces bandwidth by 90% and maintains low power consumption.
“This is a perception chip, not a computing chip, based on our original technical roadmap,” said project leader Shi Luping. “First, it balances speed and performance dynamically in a vision chip, and introduces a new computing approach that differs from existing machine vision strategies. Second, this approach mimics the dual-pathway of the human visual system, enabling efficient decision making,” Shi Luping described.
According to the team, the chip will open new directions for advances in autonomous driving and defense, as well as pave the way for the development of many new applications. Visual perception terminals in intelligent systems process huge amounts of data and must be able to handle extreme events such as unexpected driving hazards, significant lighting changes at tunnel entrances, and strong light disturbances at night.
Traditional imaging chips often fail or experience delays in these extreme situations due to limitations in power consumption and bandwidth. Tianmouc is inspired by the human visual system, the scientists added. It analyzes visual information along two pathways, one for perception and one for rapid response.
Explaining the difference with conventional chips, the scientists gave the example of flying objects that humans and Tianmouc would subconsciously avoid even if they couldn't see them clearly. Meanwhile, traditional vision chips often need to see clearly before making a decision.
The team said they developed the software, chip-based algorithms and deployed them on a Chevrolet for road testing, validating its capabilities in various extreme situations.
The chip demonstrates low latency and high performance in real-time cognitive inference across a variety of challenging environments, showing significant potential for applications in unmanned systems, according to a statement on the Tsinghua University website.
Dr. Wang Taoyi, a member of the research team, told CCTV on May 31 that Tianmouc achieved a latency of 0.1 milliseconds (ms) in autonomous driving, “which is only 1/300 of the 30ms latency encountered with traditional cameras.” Tianmouc shows significant progress compared to another brain-inspired chip, Tianjic, developed by the Tsinghua research center.
Tianjic is a chip that processes multiple AI algorithms in an energy-efficient way to control a driverless bicycle, performing tasks like turning, accelerating, and avoiding obstacles based on voice commands. Tianjic was reported in Nature in 2019.
Source: https://vov.vn/cong-nghe/tin-cong-nghe/trung-quoc-phat-trien-chip-thi-giac-nhanh-nhat-the-gioi-post1098994.vov
Comment (0)