With the rapid rise of AI applications, led by models like ChatGPT, the demand for computing power has skyrocketed. To meet the increasing computational demands of the AI era, H3C Group, under the umbrella of Tsinghua Unigroup, recently unveiled 11 new products in the H3C UniServer G6 and HPE Gen11 series at the 2023 NAVIGATE Leader Summit. These new server products create a comprehensive matrix for AI across various scenarios, providing a powerful underlying platform for handling massive data and model algorithms, and ensuring an ample supply of AI computing resources.
Diverse Product Matrix to Address Varied AI Computing Needs
As a leader in intelligent computing, H3C Group has been deeply engaged in the field of AI for many years. In 2022, H3C achieved the highest growth rate in the Chinese accelerated computing market and accumulated a total of 132 world-first rankings in the internationally renowned AI benchmark MLPerf, demonstrating its strong technical expertise and capabilities.
Leveraging an advanced computing architecture and intelligent computing power management capabilities built on the foundation of intelligent computing, H3C has developed the intelligent computing flagship H3C UniServer R5500 G6, specifically designed for large-scale model training. They have also introduced the H3C UniServer R5300 G6, a hybrid computing engine suitable for large-scale inference/training scenarios. These products further meet the diverse computing requirements in different AI scenarios, providing comprehensive AI computing coverage.
Intelligent Computing Flagship Designed for Large-scale Model Training
The H3C UniServer R5500 G6 combines strength, low power consumption, and intelligence. Compared to the previous generation, it offers three times the computational power, reducing training time by 70% for GPT-4 large-scale model training scenarios. It is applicable to various AI business scenarios, such as large-scale training, speech recognition, image classification, and machine translation.
Strength: The R5500 G6 supports up to 96 CPU cores, delivering a 150% increase in core performance. It is equipped with the new NVIDIA HGX H800 8-GPU module, providing 32 PFLOPS of computational power, resulting in a 9x improvement in large-scale model AI training speed and a 30x improvement in large-scale model AI inference performance. Additionally, with the support of PCIe 5.0 and 400G networking, users can deploy higher-performance AI computing clusters, accelerating the adoption and application of AI in enterprises.
Intelligence: The R5500 G6 supports two topology configurations, intelligently adapting to various AI application scenarios and accelerating deep learning and scientific computing applications, greatly improving GPU resource utilization. Thanks to the multi-instance GPU feature of the H800 module, a single H800 can be divided into 7 GPU instances, with the possibility of up to 56 GPU instances, each having independent computing and memory resources. This significantly enhances the flexibility of AI resources.
Low Carbon Footprint: The R5500 G6 fully supports liquid cooling, including liquid cooling for both the CPU and GPU. With a PUE (Power Usage Effectiveness) of below 1.1, it enables “cool computing” in the heat of the computational surge.
It’s worth mentioning that the R5500 G6 was recognized as one of the “Top 10 Outstanding High-Performance Servers of 2023″ in the “2023 Power Ranking for Computational Performance” upon its release.
Hybrid Computing Engine for Flexible Matching of Training and Inference Demands
The H3C UniServer R5300 G6, as the next-generation AI server, offers significant improvements in CPU and GPU specifications compared to its predecessor. It boasts outstanding performance, intelligent topology, and integrated computing and storage capabilities, making it suitable for deep learning model training, deep learning inference, and other AI application scenarios, flexibly matching training and inference computing needs.
Outstanding Performance: The R5300 G6 is compatible with the latest generation of NVIDIA enterprise-grade GPUs, providing a 4.85x performance improvement compared to the previous generation. It supports various types of AI acceleration cards, such as GPUs, DPUs, and NPUs, to meet the heterogeneous computing power requirements of AI in different scenarios, empowering the era of intelligence.
Intelligent Topology: The R5300 G6 offers five GPU topology settings, including HPC, parallel AI, serial AI, 4-card direct access, and 8-card direct access. This unprecedented flexibility greatly enhances adaptability to different user application scenarios, intelligently allocates resources, and drives efficient computing power operation.
Integrated Computing and Storage: The R5300 G6 flexibly accommodates AI acceleration cards and intelligent NICs, combining training and inference capabilities. It supports up to 10 double-width GPUs and 24 LFF (Large Form Factor) hard drive slots, enabling simultaneous training and inference on a single server and providing a cost-effective computing engine for development and testing environments. With a storage capacity of up to 400TB, it fully meets the storage space requirements of AI data.
With the AI boom surging forward, computing power is constantly being reshaped and challenged. The release of the next-generation AI servers marks another milestone in H3C Group’s commitment to “inherent intelligence” technology and its continuous drive for the evolution of intelligent computing.
Looking to the future, guided by the “Cloud-Native Intelligence” strategy, H3C Group adheres to the concept of “meticulous pragmatism, endowing the era with intelligence.” They will continue to cultivate the fertile soil of intelligent computing, explore deep-level AI application scenarios, and accelerate the arrival of an intelligent world with future-ready, adaptable computing power.
Post time: Jul-04-2023
