#USA #Supermicro #San Jose #NVIDIA #AI Performance

Leading the Way in AI Performance

Super Micro Computer, Inc. (SMCI) has set a new benchmark in the AI industry with the release of their latest NVIDIA HGX™ B200 systems. These cutting-edge systems have shown exceptional capabilities in AI performance, achieving remarkable results in the MLPerf® Inference v5.0 benchmarks. Bringing forth a combination of air-cooled and liquid-cooled systems, this innovative lineup outshines its predecessors, boasting three times the token generation per second compared to the earlier H200 systems.

Performance Highlights

The recently published results reveal that Supermicro's 4U liquid-cooled and 10U air-cooled systems topped multiple MLPerf benchmarks. Specifically designed to handle various workloads, these systems exemplify the technological strides the company has made in delivering unparalleled AI performance. The head-to-head comparison with previous generations of systems illustrates a dramatic increase in output, particularly with the Llama2-70B and Llama3.1-405B benchmarks, where Supermicro outperformed the competition significantly.

Charles Liang, the President and CEO of Supermicro, stated, "We are thrilled to lead the AI industry with our innovative technologies and systems. Our building block architecture is at the heart of our success, enabling us to roll out a diverse range of high-performance solutions."

Supermicro’s collaboration with NVIDIA is crucial to refining these systems for optimal performance. The partnership has enabled them to remain at the forefront of AI workloads, ensuring they deliver not only cutting-edge products but also industry-leading performance.

System Specifications and Features

The NVIDIA HGX™ B200 8-GPU systems offer significant enhancements in both performance and cooling technology. The introduction of new cold plates and a 250kW coolant distribution unit (CDU) marks a considerable improvement in cooling capacity within the same compact 4U form factor. This innovation allows a maximum of 64 NVIDIA Blackwell GPUs to fit comfortably in a standard 42U rack.

Moreover, the 10U air-cooled system has been redesigned to enhance thermal management while accommodating high-power GPUs. It can seamlessly integrate into existing data centers without taking up additional valuable space.

The systems have been engineered to excel at various tasks, with several benchmarks showcasing outstanding token generation per second, making them suitable for large-scale AI applications. From the Mixtral 8x7B inference tests to various Llama models, Supermicro has emerged as the top performer in these critical evaluations, providing the industry with much-needed confidence in AI system capabilities.

Industry Recognition

David Kanter, Head of MLPerf at MLCommons, commended Supermicro's achievement, emphasizing the importance of reproducible and transparent results. He noted that “the performance gains observed, especially compared to earlier generations, will certainly be appreciated by customers looking for reliable and powerful systems.”

Supermicro's comprehensive suite of over 100 GPU-optimized solutions, including their air-cooled and liquid-cooled options, allows customers unparalleled flexibility in choosing the right systems for their specific needs. The company’s commitment to optimizing total cost of ownership (TCO) while also reducing the environmental footprint aligns with the growing demand for sustainable IT solutions.

Conclusion

In conclusion, Supermicro's innovative approach to IT solutions, particularly with their latest NVIDIA HGX™ B200 systems, positions them as a leader in the rapidly evolving field of artificial intelligence. With continued advancements and a focus on quality and performance, Supermicro is paving the way for what’s next in AI technology and infrastructure. Their recent successes serve as a strong testament to their dedication to pushing boundaries and setting new industry standards.