Energy guzzler AI? Made in China: The most energy-efficient AI chips of all time

From Henrik Bork * | Translated by AI 3 min Reading Time

Related Vendors

In China, researchers have developed two extremely energy-efficient AI chips; one is optimized for speech recognition, the other for detecting epileptic seizures and has a globally leading energy consumption. These technological breakthroughs could not only change the AI landscape, but also influence the race between the US and China in the chip sector.

The energy consumption of AI chips is considered a problem, which makes the work of Chinese researchers all the more exciting.(Image: freely licensed /  Pixabay)
The energy consumption of AI chips is considered a problem, which makes the work of Chinese researchers all the more exciting.
(Image: freely licensed / Pixabay)

Chinese researchers have developed two of the most energy-efficient AI chips known to date. The first is specifically designed for offline use in speech recognition. The second is a specialty chip for the detection of epileptic seizures. Both were recently presented by the researchers at the ISSCC 2024 trade fair in San Francisco.

Chips for artificial intelligence applications usually have a very high energy demand, which makes their use in certain scenarios difficult. This is exactly what makes this recent discovery in China so interesting.

The team led by Professor Zhou Jun from the "University of Electronic Science and Technology of China" (UESTC) in Chengdu has been working on optimizing the algorithms and architecture of the chips to increase their energy efficiency, allegedly without affecting speech recognition.

Highest precision, low energy consumption.

"The chip achieves recognition energy consumption of less than two microjoules per instance, with an accuracy rate of over 95 percent in quiet scenarios and 90 percent in noisy environments, setting new global benchmarks for both energy efficiency and precision," quotes the specialist portal Interesting Engineering from the Chinese researchers.

At the ISSCC, which took place in San Francisco in February, the team showed off their new AI chip in a toy car that responds to voice commands. However, other end devices with small batteries and high requirements for voice recognition are potential applications for the new chip, such as smartwatches and devices in smart homes.

Recognize and process health signals

The second ultra-energy-efficient chip has been designed to recognize the onset signals of epileptic seizures. Installed in a wearable medical device, it can help patients by using EEG technology to warn them so that they can seek help or take other countermeasures.

This chip has an average energy consumption of about 0.07 microjoules per instance of detection. This is said to be world-leading, according to the website of the university in southwest China.

Thanks to EEG technology, this design eliminates the need for lengthy training with real patient data. Instead, the researchers optimized a "Zero-Shot" training algorithm, which allows a generally prepared AI model to deliver accurate predictions with over 98 percent precision rate, reports the South China Morning Post in Hong Kong about the new epilepsy chip from Chengdu.

This new AI chip could also be useful for other applications, such as sleep monitoring or in various brain-computer interfaces, the researchers from Sichuan write.

Competition for the USA

The team led by Professor Zhou has been working on the optimization of semiconductors for several years and maintains strategic collaborations with leading Chinese technology corporations, including SeneTime, Huawei, and electronics manufacturer BOE.

Against the backdrop of the AI race and the "Chip Wars" between the US and China, these two technological innovations once again demonstrate a diversity in the strengths of both nations.

While most experts believe the US has a few years' lead over its Chinese competition in the development and construction of the most powerful AI chips, Chinese researchers are particularly quick and successful in application-oriented research & development.

For example, GPT-4 was developed in the US, not in China. However, Chinese researchers and companies are making rapid progress in the field of LLM.

"If we can be excellent in a special business domain by training a large model with unique business data - and then connect it with many business tools in the same 'vertical' - then such a large model will not only possess intelligence, but also special knowledge," China Daily quoted Zhou Hongyi, the founder and chairman of "360 Security Technology." "It's absolutely possible that China will overtake its competitors in the U.S. in one to two years," Zhou believes. (sb)

* Henrik Bork, a long-time China correspondent for the German Süddeutsche Zeitung and the Frankfurter Rundschau, is Managing Director at Asia Waypoint, a consulting agency specializing in China based in Beijing.

Subscribe to the newsletter now

Don't Miss out on Our Best Content

By clicking on „Subscribe to Newsletter“ I agree to the processing and use of my data according to the consent form (please expand for details) and accept the Terms of Use. For more information, please see our Privacy Policy. The consent declaration relates, among other things, to the sending of editorial newsletters by email and to data matching for marketing purposes with selected advertising partners (e.g., LinkedIn, Google, Meta)

Unfold for details of your consent