Chinese Scientists Unveil Mini-Fridge-Sized AI Computer That Uses 90% Less Power Than Supercomputers
- MM24 News Desk
- Nov 1
- 4 min read
Guangdong Institute of Intelligence Science and Technology (GDIIST) unveiled the BI Explorer (BIE-1), a refrigerator-sized brain-like intelligent computer with 1,152 CPU cores that uses 90 percent less power than traditional supercomputers while achieving 100,000 tokens per second training speed and 500,000 tokens per second inference speed. The device, jointly launched by Zhuhai Hengqin Neogenint Technology and Suiren Medical Technology, runs on a household electrical socket.
Chinese scientists have unveiled what they say is the world's first brain-like intelligent computer the size of a mini fridge, which has the capabilities of a room-sized supercomputer while using 90 percent less power.
The BI Explorer computing system, or BIE-1, was revealed by members of the Guangdong Institute of Intelligence Science and Technology (GDIIST) at a forum held in the Guangdong-Macau In-depth Cooperation Zone in southern China on Friday.
The device packs the capabilities of a supercomputer – including training and inference speeds that rival traditional computing clusters – into a refrigerator-sized unit with the help of an intuitive neural network and brain-like artificial intelligence algorithm.
"It can be easily deployed in homes, small offices and even mobile environments," the institute said on its website on Friday. "It has low power consumption and low noise, and can be called a miniaturized supercomputer, making high-end intelligent computing capabilities within reach."
The device was jointly launched by Zhuhai Hengqin Neogenint Technology and Suiren (Zhuhai) Medical Technology, two companies incubated by GDIIST.
Conventional supercomputers are typically large, energy-intensive systems that can take up an entire room within a data center. These systems use a massive amount of power to not only run computing tasks but also to cool the systems.
"A traditional computing center is like a building, which requires a lot of investment in the early stages and has high energy consumption," said Nie Lei, co-director of the joint laboratory of intelligent computing systems at GDIIST.
"The [BIE-1] is only the size of a mini single-door refrigerator and can be run directly on a household socket, with power consumption of only one-tenth of traditional supercomputer equipment," Nie told China Science Daily on Saturday.
He added that the compact device would not only be suitable to be used on its own but would also provide a way to improve the energy efficiency of existing computing facilities, according to China Science Daily.
The BIE-1 integrates 1,152 central processing unit (CPU) cores – or mini processors that execute instructions and perform calculations – with 4.8 terabytes of DDR5 memory and 204 terabytes of storage space.
The device runs on an independently developed intuitive neural network, which closely mimics the brain's computational mechanisms, allowing for efficient learning, interpretable reasoning, and the ability to learn and extract patterns from small amounts of data.
WATCH ALSO: https://www.modernmechanics24.com/post/boston-dynamics-reveals-secrets-behind-humanoid-robot-grippers
The device can simultaneously process different types of information, including text, images and speech, allowing it to run model training and reasoning at high speed and efficiency, according to the institute.
The training and inference speeds of the compact computer can reach 100,000 tokens per second and 500,000 tokens per second, respectively, which the institute said were comparable to traditional computing clusters that required multiple high-end graphics processing units.
Even when performing the most complex inference tasks, the CPU temperature stays below 70 degrees Celsius (158 degrees Fahrenheit) and operates quietly.
The performance specifications suggest significant advances in thermal management and power efficiency. Traditional supercomputers require elaborate cooling systems consuming substantial additional energy. By maintaining operational temperatures below 70 degrees Celsius without specialized cooling infrastructure, the BIE-1 demonstrates that brain-inspired computing architectures can achieve computational density previously requiring massive facilities.
The brain-like approach differs fundamentally from conventional computing. While traditional systems process information sequentially through billions of transistors switching on and off, the intuitive neural network mimics how biological brains process information – with parallel processing, pattern recognition, and learning from limited data samples.
This architectural shift enables the BIE-1 to perform certain AI tasks more efficiently than systems with far greater raw computational power. The ability to extract patterns from small datasets proves particularly valuable, as training large AI models typically requires enormous datasets and energy consumption.
According to the institute, the BIE-1 could be used in many different industries, such as home health monitoring, customized tutoring for children and personalized AI help in offices.
The device could also offer non-professional users access to brain-inspired computing technology and allow for intelligent computing to become available in even more situations and environments.
The compact form factor and household power requirements open possibilities previously impractical. Small medical clinics could run sophisticated diagnostic AI locally. Schools could provide students with access to powerful computational resources without building data centers. Remote research stations could deploy advanced computing capabilities without infrastructure investments.
Whether the BIE-1 can deliver on these promises at scale remains to be proven. Real-world performance across diverse applications, long-term reliability, software ecosystem development, and competitive pricing all need demonstration before widespread adoption becomes feasible.
But if GDIIST's claims hold up under independent testing, this mini-fridge-sized computer could represent a genuine paradigm shift in how we think about deploying advanced AI computational resources.
Comments