Synced (a Chinese information platform) reported that Cambricon Technology (寒武纪科技), a leading Chinese smart chip firm, will launch its second generation of AI chip, MLU270, for cloud data center servers in this May. One year ago, Cambricon launched the first-generation product, MLU100, and obtained purchase orders from Inspur, Lenovo, Sugon, DiDi and HIKVISION.
MLU270 was developed successfully at the beginning of this year, and its main application scenario is cloud computing for algorithm training. MLU270 upgrades its chip architecture from MLUv01 to MLUv02. Considered that explosive growth of video data volume has occupied plenty of computing source for data center, Cambricon builds video decoding unit in MLU270, which aims at the video processing market.
In terms of performance parameter, MLU270 seems like a direct competitor of Nvidia’s Tesla T4. MLU270 supports INT16, INT8 and INT4, whose peak performance are 64 Tops, 128 Tops and 256 Tops. Compared with MLU270, Tesla T4 supports FP16, INT8 and INT4, whose peak performance are 65 Tops, 130 Tops and 260 Tops. As for power consumption, they are around 75w.
Cambricon emphasizes that MLU270 has advantage on integer calculation performance, because Cambricon makes a breakthrough on low-precision training algorithms. MLU270 seems like to be promoted by this unique tech, rather than 7nm technology. Compared with MLU270, Nvidia’s Tesla T4 GPUs adopts 14nm technology, while AMD’s Radeon Instinct MI60 GPUs and MI50 GPUs adopt 7nm.
The chips for cloud data center server market has been almost dominated by several global technology companies at present, including Intel, Nvidia, AMD, Qualcomm, etc. So how dare of Cambricon to enter this market and compete with these giants?
Born from Institute of Computing Technology Chinese Academy of Sciences, Cambricon has advantages on R&D of AI chips. CHEN Tianshi (陈天石), the founder and CEO of Cambricon, is an outstanding expert in the field of processor architecture and AI with high reputation all over the world. He obtained received his Ph.D. in engineering from the School of Computer Science, China University of Science and Technology in 2010. He has attracted lots of scientists to join his firm, who have an average of more than 9-year R&D experience.
In 2016, Cambricon launched Cambricon-1A processor, which is the world's first commercial deep learning dedicated processor for smartphones, security surveillance, wearables, drones and smart driving. In 2018, Cambricon launched Cambricon 1M chip for edge computing, and the MLU100, the first in a new chip series for cloud computing.
Cambricon has grown to be a unicorn company valuated USD 2.5 USD after its B series funding on June, 2018. It has financed for 6 rounds, and its main investors includes many state-owned large funds, such as CMCDI, China Venture Capital Fund, Chinese Academy of Sciences, etc. Last year, CHEN Tianshi revealed that Cambricon might list in A-share market in the future.
According to Allied Market Research, the global machine learning chip market was about USD 2.4 billion in 2017. It is estimated that this market will reach about USD 37.8 billion by 2025, with a CAGR of 40.8%. Cambricon has great opportunity to gain much market share and grow to be a large company in the field of AI chips.