According to the news on 19 September, Huawei's Vice Chairman and Chairman-in-Office Xu Zhijun said publicly today that not every company has to train its own basic big model.
We have to face the reality that the US sanctions against China in the field of AI chips will not be lifted for a long time, and China's semiconductor manufacturing process will be in a backward state for quite a long time as it is also subject to the US sanctions, which means that the sophistication of the chips we can make will be constrained. This is a challenge we must face in building arithmetic solutions.’ Xu Zhijun said.
Based in China, only arithmetic built on actually accessible chip manufacturing processes is sustainable in the long term.
In Xu Zhijun's view, Huawei's strategic core is to fully seize the opportunity of the change of artificial intelligence, based on the actual available chip manufacturing process, computing, storage and network technology co-innovation, pioneering computing architecture, to create a ‘super node + cluster’ system arithmetic solutions, long-term sustainable to meet the arithmetic demand.
From Huawei Pangu's practice in the industry, the one-billion parameter model can meet the needs of scientific computing, prediction and decision-making business scenarios, such as rainfall prediction, drug molecule optimisation, and process parameter prediction, and the one-billion parameter model has been widely used in PCs, mobile phones, and other end-side devices.
And ten billion parameter models can meet the needs of a large number of domain-specific scenarios oriented towards NLP, CV, and multimodality, such as knowledge quiz, code generation, seat assistant, and security detection. Complex tasks oriented towards NLP and multimodality can be accomplished with a hundred billion parameter model.
