Algorithms, data, and computing power are the three key factors driving the rapid development of AI deep learning technology. In order to support stronger generalization ability, higher prediction accuracy, and the algorithm model becomes more and more complex, more and more data needs to be labeled and processed in time, and computing performance becomes the key. Recently, the AI server project launched by ODCC has ushered in new progress. The super AI computing platform X-MAN3.0 developed by Baidu experts will be unveiled at the summit! The AI computing platform provides ultimate computing performance, supports large and complex algorithm models, and can process massive amounts of data quickly and in a timely manner.
3 key technical challenges solved by X-MAN
In order to provide more computing performance and optimal computing efficiency, X-MAN faces three key technology challenges in system design: how to effectively improve stand-alone computing performance, how to achieve high scalability of multi-machine acceleration, and how to balance The matching relationship between the CPU and the AI acceleration chip to avoid system bottlenecks.
X-MAN has created 6 industry firsts, leading the industry technology trends
With a deep understanding of the AI business model, solid hardware infrastructure technology and accurate prediction of industry trends, Baidu designed the X-MAN AI computing platform series and created six industry firsts. Among them, X-MAN1.0 released in Q2 2016 created five industry firsts: the first implementation of single-machine expansion of 16 AI accelerator cards; the first realization of 4/8/16/32/64 card system-level elastic expansion; the first realization of AI Computational hardware decoupling architecture (Note: Facebook's Q1 released in Q1 2017, Microsoft's HGX-1 released in Q1 2017 have adopted a similar design concept); the first implementation of PCIe Fabric technology architecture, support dynamic allocation of AI calculations according to business needs Resources; for the first time, virtual machine multi-card high-speed lossless communication technology (native performance). The X-MAN2.0 industry released in Q3 2017 applied the cold plate liquid cooling technology to AI computing products, effectively solving the heat dissipation challenge of high-performance AI accelerator cards. (Note: Google 2018 Q2 released TPU3.0 also A similar technique was used).
X-MAN3.0 accelerates the diversification of AI chips
X-MAN3.0 adopts the design idea of modularization of hardware system, standardization of interconnection interface and high flexibility of interconnection topology, which not only improves the compatibility of diversified AI acceleration chips, but also promotes the ecological health, fast and sustainable of AI acceleration chips. development of. The X-MAN 3.0 features three key technical features, including 48V power supply technology, standardized AI accelerator card interface definitions, and flexible interconnect topologies.
X-MAN large-scale deployment, helping Baidu AI strategy to quickly land
With powerful stand-alone computing performance, industry-leading multi-machine scalability, and flexible AI computing resource dynamic allocation and adjustment capabilities, X-MAN series products have been deployed on a large scale in Baidu, widely used in speech recognition, computer vision, natural speech. The deep learning model training of basic AI technology such as handling and unmanned vehicles helps the rapid development of Baidu's AI strategy.