At the end of 2022, The emergence of generative AI, marked by ChatGPT, ignited the market and captured industry-wide attention towards large AI models. This signalled a pivotal shift in AI development, with large models becoming a key driver of the new technological revolution and industrial transformation. The intensifying "Battle of the Models" led to an enormous demand for AI computing power, while the localisation of computing power has become increasingly urgent.
Large Model Training: A High-Performance Computing Application
Deploying computing resources using HPC architecture is an ideal solution for large-scale AI model training, effectively meeting the demands for high-end computing power, data storage, and processing capabilities. As a leading provider of high-performance computing solutions in China, Lianzhi Technology is committed to collaborating with multiple partners to construct HPC architectures suitable for large model training and inference. This initiative aims to build a comprehensive converged computing power management platform enabling centralised management, efficient utilisation, and multi-faceted coordination of heterogeneous computing resources.
ClusterTech Beijing Limited CHESS X Hygon Heterogeneous Acceleration Card
After the joint testing with Hygon AI Acceleration Card series, ClusterTech Beijing Limited’s HPC cluster management platform, CHESS has proved to install successfully on both Hygon heterogeneous acceleration cards and Hygon CPU environments due to its exceptional algorithm and application optimisation capabilities. It also demonstrated excellent compatibility with Zhongke Controllable X73, X74, and X78 series heterogeneous acceleration servers, operating stably, reliably, and with high performance to meet complex computational demands.These results fully validates that CHESS possesses high integrity, openness, and compatibility, adapting well to domestic GPU servers and system platforms, fulfilling the requirements for developing large domestic models.
The Hygon heterogeneous acceleration card is designed specifically for AI and deep learning. It comprehensively supports deep learning training scenarios, effortlessly handles complex neural network training, providing robust computing power for artificial intelligence applications. It already enables full-scale application of major models including LLaMa, GPT, Bloom, ChatGLM, Wudao, and Zidong Taichu, and is fully compatible with domestic large models such as ERNIE Bot (Wenxin Yiyan), with extensive applications in key sectors such as telecommunications, finance, education, research, and AI.
As a key participant in the Hygon ecosystem, ClusterTech Beijing Limited will combine the computing power of Hygon Heterogeneous Acceleration Cards with its strengths in supercomputing algorithm and application expertise, deploying the independently controllable CHESS platform. CHESS provides centralised management of heterogeneous computing resources, schedules diverse tasks, supports complex task orchestration, integrates multiple AI application frameworks to optimise application performance, thereby quickly establishing a stable, reliable, and user-friendly heterogeneous cluster system.
Looking ahead, computing power will inevitably grow more intelligent. ClusterTech Beijing Limited will continue to focus on domestic GPU computing power and large model development needs. It aims to build a mature, heterogeneous converged computing ecosystem, delivering more comprehensive and efficient parallel computing capabilities for large model training and complex applications, thereby overcoming the computational bottlenecks hindering large model advancement.
*ClusterTech Beijing Limited is the Beijing subsidiary of ClusterTech. Learn more: http://www.lianzhitech.cn/
