Inspur Showcases AI Computing Solution Based on GPU

At GTC China 2017, held in Beijing on Sept. 25-27, Inspur showcased its full line of advanced NVIDIA GPU-based AI computing products and solutions. Inspur became the first manufacturer to deliver AI supercomputers NF5288M5 (AGX-2) based on NVIDIA® Tesla® V100 to Chinese AI customers.

The NF5288M5 (AGX-2) server showcased on this event is the first high-density and high-performance supercomputing server in the industry that supports Tesla®V100 and NVLink 2.0. The server integrates 8 most advanced GPUs at present with the High-speed Interconnected technology within a space of 2U and has good I/O scalability, especially developed for AI computing and application service providers demanding higher computing density in the same physical space. Currently, the AGX-2 server based on Tesla®V100 has undergone several AI model training tests on the platforms of China's tech giants Tencent, Alibaba and Baidu.

In addition to joint innovation at the product level, Inspur is working together with NVIDIA also in continuously perfecting the GPU ecosystem. With the foundation of the Inspur-NVIDIA Cloud Supercomputing & Application Innovation Center, Inspur and NVIDIA are co-operating in promotion of GPU-related technologies in the AI field for creating its new values and industry-scale reconstruction.

In Caffe-MPI 2.0, the latest cluster edition of open-source deep learning framework issued by Inspur, the updated editions of cuDNN and NCCL are adopted. When the Resnet model is applied to the 16-card clusters, the training performance achieved a 15 times higher acceleration ratio to that of the single-card training.

Inspur considers talent cultivation as a key aspect of AI ecosystem development and cooperates with NVIDIA in nurturing AI computing talents, jointly establishing the ICCE GPU Engineer Certification Training Program which has become the first and the largest specialized training program for GPU applications. The training program covers theoretical explanations on the latest GPU structure, CUDA programming & optimization and study of industrial cases, integrates theoretical knowledge with practical applications and has won high reputations among the trainees since its establishment. So far, ICCE certification trainings have been successfully held for several years and produced nearly 1,000 certified trainees.