re <<How about>> ... energy
scmp.com
China’s analogue AI chip runs 12 times as fast on 1/200th the energy of digital rivals
Chinese scientists go back to the future for inspiration that could reshape the power-hungry artificial intelligence model
Ling Xinin Ohio
Published: 2:00pm, 23 Jan 2026Updated: 5:31pm, 23 Jan 2026
A radically different kind of chip created by Chinese researchers can now handle real-world data tasks, potentially reshaping artificial intelligence systems’ reliance on power-hungry digital processors, its developers said.
Building on work reported in October, the Peking University team’s ultra-fast, energy-efficient analogue chip has moved beyond solving basic mathematical problems and can now power applications such as personalised recommendation and image processing.
In a paper published on Monday by the journal Nature Communications, lead author Sun Zhong and his colleagues said the chip achieved a 12-fold speed increase over advanced digital processors, while improving energy efficiency by more than 200 times.
These results were based on the training of recommendation systems using data sets that were comparable in scale to those of Netflix and Yahoo, according to the peer-reviewed paper.
In image-compression tests, the system reconstructed images with almost the same visual quality as full-precision digital computing, while cutting storage requirements by half, the researchers wrote.
In a social media post, Sun wrote that the study “pushes the boundary of analogue computing one step further”. He added that the new chip had handled more complex tasks while retaining the speed and energy advantages of analogue computing.
A reviewer of the paper said the experimental results – particularly the “orders-of-magnitude improvements in speed and energy efficiency” – when compared with conventional digital chips, showed the technology’s potential for industrial applications.
The idea behind the chip, which performs calculations using physical signals rather than digital code, is not new. Scientists were exploring analogue computers decades ago, until the technology was sidelined as digital chips became faster, cheaper and more reliable.
Unlike digital computers, which process information step-by-step using binary code made of zeros and ones, analogue computing represents numbers as continuously changing physical signals, such as electrical currents or voltages.
In theory, this allows many calculations to happen at the same time, offering huge gains in speed and energy efficiency for certain tasks.
With AI systems heavily reliant on power-hungry digital processors, the cost of moving data between memory and computing units has become a major bottleneck.
Advances in materials, circuit design and algorithms have renewed interest in analogue hardware, which performs calculations directly where data is stored.
In a paper published last year, Sun’s team showed that analogue chips could dramatically accelerate basic mathematical operations – up to 1,000 times faster than top digital processors, such as the Nvidia H100 graphics processing unit – while consuming far less energy.
In the latest study, the researchers adopted a technique known as non-negative matrix factorisation (NMF), a powerful tool for data reduction that extracts underlying structures from vast and complex information, such as user behaviour or image pixels, Sun told Science Daily on Thursday.
The technique has been widely used in image analysis, data clustering and personalised recommendation systems, but as data sets grow to millions of entries, traditional digital hardware increasingly struggles with computational complexity and memory bottlenecks, he said.
To address these limits, the team built an analogue computing chip based on resistive memory and redesigned its core circuitry to carry out the most demanding part of the algorithm in a single step.
According to the Science Daily report, the researchers were able to achieve the same results with fewer computing units, significantly reducing the chip’s size and energy consumption.
“Our study opened a new path for solving complex data problems in real time and highlighted the huge potential of analogue computing for practical applications,” Sun said in the interview.
In a social media post, Sun noted that he had “loved” the NMF technique since it was first proposed in 1999 by two Korean-American scientists in a paper published by the journal Nature.
“It’s deeply rewarding to see it now brought into the realm of in-memory analogue computing, 27 years later,” Sun wrote. |
