Today, Huawei held the launch conference of the Ascend 910 AI processor and MindSpore computing framework in Shenzhen. At the press conference, Huawei said this is the industry’s fastest performing AI processor.
In October 2018, at the Huawei 2018 General Assembly, Xu Zhijun, the rotating chairman of Huawei, first elaborated on the AI strategy and officially announced two AI chips, the Ascend 910 and the Ascend 310. Xu Zhijun said that the Ascend 910 is a chip with the highest density calculation on a single chip. It comes with the 7nm + EUV process, 32-core self-developed DaVinci architecture, half-precision (FP16) computing power of 256 Tera-FLOPS, integer precision (INT8) computing power of 512 Tera-OPS.
According to Xu Zhijun, Ascend 910 has two highlights:
First, it is the world’s most powerful and fastest training AI chip. Its computing power is twice that of the international top AI chips. It exceeds the latest and strongest CPU by 50 times. Its training speed is also increased by 50%-100% in comparison to the latest and most recent strong chips.
Second, the new generation of AI open-source computing framework MindSpore is even more powerful. Its innovative programming paradigm is easier to use by AI scientists and engineers. The computing framework can meet the needs of the terminal, edge computing, and cloud computing. It better protects data privacy. This chip provides the best performance due to the AI chip, maximizing the use of chip computing power.
The test results show that the Ascend 910 fully meets the design specifications in terms of computing power. What is important, the power consumption required to reach the specification calculation power is only 310W, which is significantly lower than the 350W of the chip. Xu Zhijun said: ‘Huawei has used the Ascend 910 for actual AI training tasks. This chip and MindSpore show a nearly double performance improvement compared with the existing mainstream training single card TensorFlow. In the future, Huawei will continue to invest and introduce more AI processors to continue providing more abundant, more economical and more suitable AI computing power.’
Huawei also released the full-scenario AI computing framework MindSpore today. Last year, at the Huawei Full Connect Conference, the company announced the AI framework should be development-friendly, efficient, and more important to adapt to each scenario, including endpoints, edges, and clouds. After nearly a year of hard work, MindSpore has made significant progress in all three areas. The MindSpore architecture supports large and small cores and adapts to the full deployment of the entire scene. MindSpore achieves cross-scenario collaboration under the premise of ensuring that user privacy data is protected by coordinating processed gradients.