The Yunfeng T10 will be available in the first quarter of 2020.
2019 is the year that the industry recognizes the commercialization of artificial intelligence. AI chips are one of the key elements among them, and they may become a billion-yuan market in the near future.
The “Ebara Technology” recently contacted was established in March 2018. The goal is to provide customers with domestic general-purpose artificial intelligence solutions for cloud training and inference, serving the financial, medical, education, transportation and other industries, and for the public cloud. , AI, private cloud and hybrid cloud provide AI computing power.
In general, cloud data center artificial intelligence training will require training chips with high computing power, low power consumption, strong interconnection, and support multiple training algorithms to meet their requirements for versatility and energy consumption.
In response to these requirements, Ebara Technology recently released its first artificial intelligence training product, the “Yunya T10”. According to the introduction, this is an artificial intelligence training acceleration card and a dual-slot target card for cloud data centers. It supports PCIe 4.0. It will be available in the first quarter of 2020 and mass production. The 25GB bi-directional backplane interconnection solution is different from the industry-wide InfiniBand networking, which can greatly reduce the complexity and cost of networking with the same interconnection bandwidth. Its single card single precision (FP32) computing power reaches 20TFLOPS, and supports mixed precision calculation of single precision FP32 and half precision BF16.
In addition, based on the design concept of reconfigurable chips, the computing core of the “Sisi” chip released by Ebara Technology includes 32 general-purpose scalable neuron processors (SIPs), each 8 SIPs are combined into one Scalable Intelligent Computing Group (SIC). SIC achieves high-speed interconnection through HBM, and through on-chip scheduling algorithms, data is calculated during relocation to maximize SIP utilization. The chip is built on a 12nm FinFET process with a total of 14.1 billion transistors, which overcomes the 2.5D packaging problem.
At the same time, Ebara Technology also provides a computing and programming platform, “YuShu”, which supports mainstream deep learning frameworks, provides a complete compilation, debugging, and tuning tool chain, and opens the SDK at the hardware layer to provide in-depth developers Granular computing power programming interface. In addition, the company also launched the “Cloud Rage T11” based on OCP Acceleration Module (OAM), which has higher performance than PCIe standard cards. Ebara Technology claims that they are the first company in China to make an OAM module.
Talking about the pain points of the development of China’s AI industry, Zhao Lidong, CEO of Ebara Technology, told us that first of all, the previous GPU was mainly used for game acceleration and image rendering, which could not support precision, computing power and architecture. The diversified needs of the artificial intelligence industry; secondly, cloud training is the source of the entire architecture, and a training card is often easy5US $ 000-8000; Third, AI chips have long been monopolized by foreign giants, closed and not open sourced. When customers in the industry make customized and differentiated products, they rely on engineers from foreign giants, and their products must be upgraded. Depending on the pace of chip supplier upgrades, the localized technical support of foreign companies often cannot meet the differentiated needs of the domestic market.
Liaoyuan Technology announced in June this year that it has received RMB 300 million in Series A financing, led by Red Dot Ventures China Fund, Haisong Capital, Cloud and Capital, Tencent Investment, Sunshine Capital, Xinzhongli Capital Follow-up investment; announced in July 2018 that it had obtained a pre-A round of financing of RMB 340 million, led by Tencent, and also continued to follow-up with Capital (Funds of Wu Yuefeng Capital), Zhenge Fund, Data Capital, Cloud and Capital ; In April 2018, obtained the angel round / seed round financing. The investors include Yihe Capital, Zhenge Fund, Data Capital, Cloud and Capital.
It is worth noting that Tencent, which does not easily invest in hardware companies, led the AI chip company in June last year as a major shareholder. Zhao Lidong said that they have worked closely with Tencent’s project on general artificial intelligence application scenarios and will expand more scenarios in the future.
Zhao Lidong mentioned that they currently have three main business directions. The first is cloud service providers, including public cloud, private cloud, and hybrid cloud service providers. The second is industry service providers, such as financial services, insurance, In the medical and transportation industries, investors in these industries can help Liaoyuan Technology quickly land in vertical fields; the third is smart cities and AI supercomputing centers.
“In all three directions, we hope to find a strategic partner for in-depth cooperation, and eventually change from one point to many points, from the point to the line, from the line to form a surface.” Zhao Lidong said.
Speaking of the mass production capacity of the Yunyao T10, Zhang Yalin told us that in general, the large-scale mass production of chips means that the yield is in line with requirements and the supply is not a problem. The latter depends on the customer’s ability to expand. Now Ebara ’s “mass production yield fully meets the GlobalFoundries yield index … It usually takes one year for a chip to go from mass production to mass production. This year, it is necessary to climb the yield and stabilize it. Performance, reliability, and various characterization tests to ensure that the chip can pass tests such as extreme temperature and aging. ”
Speaking of future R & D planning, the company’s founder and COO Zhang Yalin said that Although the company currently mainly develops cloud training chips, it will definitely deploy cloud reasoning chips in the future. Compared to training chips, the inference chip is not as versatile. It needs to be optimized more powerfully in the field of certain business models in order to truly provide effectiveness and be more sensitive to power consumption and cost.
Zhao Lidong said that Liaoyuan Technology’s current goal is to build a top-level engineering team, complete product research and development and mass production, and realize the hot start of products.