
When DeepSeek, a large domestic model, exploded around the world with its “low cost and high IQ”, PetaIO and BAIXIN delivered a hardcore answer with an all-domesticized arithmetic base - the BAIXIN servers equipped with PetaIO SSDs increased the overall performance of the DeepSeek model by more than 20%, and the single computer power crushed 30 traditional servers! BAIXIN servers equipped with PetaIO SSDs have improved the comprehensive performance of DeepSeek models by more than 20%, and their single-computer arithmetic power has crushed that of 30 traditional servers! This is not only the deep integration of AI and storage technology, but also a milestone breakthrough in localized substitution.
BAIXIN's self-developed Hengshan 326TA homegrown server has successfully cracked the problem of highly concurrent storage and dynamic load balancing in DeepSeek-V2 model training through the synergistic optimization of heterogeneous access and storage technology and PetaIO:
• Arithmetic density jumps:The performance of a single machine is equivalent to 30 traditional servers, supports training of hundreds of billions of parametric models, and the failure rate is 40% lower than that of imported solutions;
• Leap in reasoning efficiency:
Based on the hardware tuning capability of the Rise architecture, combined with the high throughput characteristics of PetaIO SSDs, the model inference latency has been reduced by 15%, and millisecond response has been realized in the real-world testing of the government cloud;
• Energy management optimization:
Through intelligent resource scheduling algorithms, energy consumption per unit of arithmetic is reduced by 25%, helping the application performance of enterprise AI to run better.
As the core storage component of BAIXIN server, PetaIO supports DeepSeek full-scenario applications with three major advantages:
1. Highly concurrent storage: supporting millions of IOPS per second to meet the massive data throughput demand for training of hundreds of billions of parameter models;
2. Dynamic load balancing: through intelligent caching algorithms, the storage bandwidth utilization rate is increased to 95%, avoiding arithmetic idling;
3. Localization of the whole link: from the main control chip to the flash particles, it is 100% independently controllable and assisted in the passage of the national Xinxin technology for the BAIXIN self-development. Work Committee certification, providing a secure base for sensitive scenarios such as finance and medical care.
•Government Cloud: Supporting Hefei's “Chaohu Bright Moon” arithmetic cluster, realizing the DeepSeek-R1 model's second-level decision-making response in smart city governance;
•Intelligent Manufacturing: In the automotive R&D scenario, Baxin Server improves DeepSeek's simulation computation efficiency by 30%, helping automotive companies shorten their R&D cycle;
•Biomedical: Relying on the high-speed data read/write capability of PetaIO SSD, the efficiency of AI drug screening is 5 times higher than that of traditional solutions.
At present, Baxin has successfully completed the adaptation and performance tuning of DeepSeek-V2 large model, marking a milestone breakthrough in the field of large model of domestic AI computing power platform.
PetaIO and BAIXIN 2024 signed a strategic cooperation agreement at the “Shanxi Xintron Ecological Development Roundtable” to jointly promote the construction of the Xintron industrial ecosystem. The cooperation focuses on the in-depth integration of localized storage technology and server hardware, with the goal of providing an autonomous and controllable arithmetic base for AI, cloud computing and other scenarios. The results of the cooperation between the two sides have landed in Hefei's “Chaohu Bright Moon” arithmetic cluster, which supports the real-time decision-making of DeepSeek-R1 model in the governance of smart cities. In the government cloud scenario, PetaIO's encryption technology (in line with the certification of the National Information and Innovation Industry Committee) guarantees the security of sensitive data, and the failure rate is 40% lower than that of imported solutions.