Chinese "AI refrigerator" sits under the desk
The computer Neogenint BIE-1 is intended to process large AI models, but also to be compact, quiet, and economical enough for offices.
AI workstation Neogenint BIE-1 under a desk.
(Image: Guangdong Institute of Intelligence Science and Technology (GDIIST))
Chinese manufacturer Neogenint, in cooperation with a research institute, has developed the relatively compact BIE-1 computer for special AI algorithms. The BI Explorer-1 (BIE-1) is designed for "Brain-like Intelligence". It was presented at an event by the Guangdong Institute of Intelligence Science and Technology (GDIIST), which cooperates with Neogenint, among others.
Videos by heise
According to GDIIST and Neogenint, the BIE-1 delivers very high AI computing power with particular efficiency. Its compact design allows it to fit under a desk, it is said to work with a standard power outlet, and to run quietly enough for office environments.
The development partners promise up to 12.6 million tokens/s when inferencing an LLM with the so-called Intuitive Neural Network (INN) on the BIE-1. This "brain-like" INN appears to be a development of the GDIIST, which also researches medical AI systems.
(Image:Â Shanghai Neogenint Intelligent Technology Co.,Ltd.)
The BIE-1 consists of three modules, each containing 384 processor cores, as well as 1.5 TByte RAM, three "GPUs" with 24 GByte local memory each, one M.2 SSD with 4 TByte, and three 16-TByte hard drives. The three modules together provide 1152 CPU cores, 4.5 Terabytes of RAM, 12 TByte of flash storage, and 144 TByte on hard drives.
According to Neogenint, a single module delivers 7000 tokens/s with an INN thread. With 600 threads, it should be 4.2 million tokens/s (7000 Ă— 600), and with three modules (and consequently 1800 threads), the mentioned 12.6 million tokens/s. The power consumption of the entire unit is said to be a maximum of 2000 watts.
Neogenint is currently not providing prices or delivery dates. Neogenint also does not disclose how many physical processors from which manufacturer are used (2 Ă— 192 cores or 1 Ă— 384 cores) and which type of GPU accelerator. GDIIST also generally cooperates with the Chinese AI chip developer Cambricon.
The website of the company Shanghai Neogenint Intelligent Technology does not use an encrypted connection, therefore the source is not provided as a link here: http://www.neogenint.com/bie
(ciw)
