The AGX-2 was unveiled at Inspur booth on May 10. The AGX-2 supports up to 8 lastest NVIDIA Tesla V100 GPUs, offering either PCI-e interface or NVLink 2.0 for faster interlink connections between the GPUs, reaching peak performance results of up to 150GB/s. AGX-2 provides great I/O expansion capabilities, supporting 8x NVMe/SAS/SATA hot swap hard drives and high-speed cluster interconnect for up to 4x 100Gbps EDR InfiniBand connector cards. The unique air cooling or air-liquid hybrid cooling design enables deployment of Green Datacenters with lower PUE. AGX-2 supports both air-cooling and on-chip liquid-cooling to optimize and improve power efficiency and performance.
According to the LINPACK benchmark results, the AGX-2 achieves 29.33 TFLOPS, which is 2.47 times faster than the testing on NF5288M4 manufactured by Inspur with 4 GPUs in a 2U form factor. Regarding the real performance of training on an AI model, the AGX-2 delivers 1165 images/s, which is 2.49 times faster than the NF5288M4 with 4 Tesla M40, when the GoogLeNet model is trained with TensorFlow.
Leijun Hu, Vice President of Inspur Group, stated: "NVIDIA is the world leader in visual computing and is reshaping the next era of AI computing. Inspur partners with NVIDIA to announce the new and innovative AGX-2 GPU server today, which offers high computing density and enables faster, easier multi-GPU computing. The cooperation between the two companies also shows Inspur's capability to develop high performance computing servers to propel AI, deep learning and advanced analytics, and we are hoping to provide even more energy-efficient computing solutions to serve customers around the world."
"Inspur , while having a long term cooperation with NVIDIA, has rich R&D and practice experience in computing system for deep learning as well as long time history cooperation with NVIDIA", stated Marc Hamilton. VP, Solutions Architecture and Engineering at NVIDIA. "The launch of AGX-2, the Ultra Dense Server which employs world's top NVIDIA Tesla P100 GPU and high-speed interconnect NVLink technology, will comprehensively increase efficiency in AI and scientific engineering computing in terms of performance and energy consumption, and and it will provide both Chinese and global enterprises with leading high-performance computing capability."
Inspur keeps track of all the emerging intelligent computing technologies and applications, and is able to provide customers with the most complete GPU server product solutions that can support 2\4\8 GPU accelerators for a standalone system. Inspur and Baidu are also working together to develop an extendable SR-AI Rack Scale System designed to support 16 NVIDIA Tesla GPU Accelerators in a box. This powerful system can immediately shorten data processing time, visualize more data, accelerate deep learning frameworks, and design more sophisticated neural networks.
Inspur has been working with partners and customers around the world to help them accelerate solutions, automate operations, and gather better insights to make smarter decisions. In China, Inspur accounts for more than 60% of AI computing server market, and works closely with leading AI companies, such as Baidu, Alibaba, Tencent, iFLYTEK, Qihoo 360, Sogou, Toutiao, and Face++, to achieve significant performance improvements for voice, image, video, search, networking and other applications.