On January 21, Chinese AI company Cambricon's MLU 290 and accelerator card, Xuansi 1000 smart accelerator debuted after mass production.
The MLU 290 smart chip is Cambricon's first training chip, using TSMC's 7nm process, integrating 46 billion transistors. It supports AI training, inference, or hybrid AI compute acceleration tasks.
Cambricon is the first AI chip stock in China. Previously Cambricon has launched MLU 100 and MLU 270 in the cloud, both on a 16nm process.
MLU 290 is fully upgraded in terms of architecture, memory, and interface. Compared to MLU 270, MLU 290 achieves a 4x increase in peak arithmetic power, 12x increase in memory bandwidth, and 19x increase in inter-chip communication bandwidth.
Based on MLU 290, Cambricon has also launched MLU290-M5 intelligent accelerator card and Xuansi 1000 intelligent accelerator.
The accelerator card adopts Cambricon's self-developed MLU-Link technology, providing AI arithmetic power up to 1024 TOPS with maximum heat dissipation of 350W; the Xuansi 1000 intelligent accelerator integrates four MLU 290s, with a maximum AI arithmetic power of over 4.1 PetaOPS INT4.
It is worth noting that Huawei and NVidia have long launched the same type of chip with a 7nm process. After losing Huawei as a major customer, Cambricon's revenue in the first three quarters of last year was RMB 158 million, only 35% of 2019, with a loss of more than RMB 300 million.
However, the launch of the advanced process chip has stimulated capital markets, with Cambricon up about 18% today.
Cambricon currently has a market cap of RMB 68 billion, although it is still below the highest market cap of RMB 100 billion reached after the company's IPO.