Cambricon's first AI training chip completes mass production
Cambricon's AI training chip, the Siyuan 290 smart chip. [cambricon.com]
Cambricon in the Zhongguancun Demonstration Zone announced on Jan 22 that the Siyuan 290 smart chip and accelerator card and the Xuansi 1000 smart accelerator have officially debuted after mass production.
The Siyuan 290 smart chip is Cambricon's first training chip. It adopts 7nm advanced process technology, integrates 46 billion transistors, supports MLUv02 extended architecture, and fully supports AI training, reasoning or hybrid artificial intelligence computing acceleration tasks.
The Cambricon Siyuan 290 chip and accelerator card have been adapted to some hardware, and large-scale shipments of the product are now underway.
A Xuansi 1000 computing unit can replace a small supercomputing center
Cambricon's first smart accelerator, Xuansi 1000, contains four Siyuan 290 smart accelerator cards, with a maximum AI computing power of more than 4.1 PetaOPS INT4. A Xuansi 1000 computing unit is enough to replace a small traditional supercomputing center.
In recent years, the complexity of AI algorithm models has grown rapidly, and higher requirements have been placed on computing power and training speed. To build a more powerful computing platform, interconnection technology among multiple chips has become a rigid demand in the market.
The next-generation AI computing center requires more intelligent chips to work seamlessly while also maintaining high computing efficiency, thereby providing massive computing power to meet the needs of ultra-large-scale training.
In this regard, Cambricon rethinks the architecture of the future AIDC, using the unified MLU-Link multi-core interconnection technology for communication inside and outside the Xuansi 1000 smart accelerator. This allows the Siyuan 290 smart chip's interconnection range to be expanded from a single machine to the POD and even the entire computing center.
In addition to the standard configuration of the POD, if the computing center conditions permit, through the MLU-LinkTM multi-core interconnection technology, 1,024 or more Siyuan 290 interconnects can be realized, and no additional network card is required to achieve seamless acceleration.
Cambricon was listed on the Science and Technology Innovation Board last year as the first domestic AI chip stock.
In the future, Cambricon will continue to develop and productize smart chips and basic system software for diversified AI application scenarios and needs and provide underlying computing power support for the broad application of AI technology in various industries.