In a significant announcement during the SC26 convention, MiTAC Computing Technology Corporation, a leading provider in high-performance server solutions and a subsidiary of MiTAC Holdings Corporation, showcased its innovative full-stack infrastructure for agentic AI deployment. The theme of the exhibit, "Scaling Agentic AI. Built on Full-Stack Infrastructure," focused on how a comprehensive approach is necessary to tackle the computational bottlenecks associated with this advanced technology.
The Need for Unified Infrastructure
Rick Hwang, the president of MiTAC Computing, emphasized that to scale agentic AI effectively, companies must adopt an integrated infrastructure rather than relying on fragmented systems. This holistic setup is vital for creating sustainable AI factory projects, allowing for a more seamless integration of different components needed for efficient AI operation.
Advanced Architecture for Inference
A notable collaboration with Skymizer is paving the way for a groundbreaking inference architecture at MiTAC. This setup designates the computationally intensive task of pre-filling data to GPUs while allocating the memory-heavy decoding tasks to Skymizer's HTX301 accelerator cards. This strategic division enables businesses to tailor each processing step according to their specific workload demands, ensuring high efficiency and processing speed while keeping model inference on-site.
For the pre-filling stage, MiTAC offers the G8825Z5 server—equipped with powerful AMD EPYC™ processors and AMD Instinct™ GPUs—to handle the computationally heavy lifting and transfer critical data effectively to the decoding server via a high-speed network. On the decoding front, the air-cooled G4520G6 server is designed for enterprise-level AI workloads, capable of executing large-scale models like the astronomical Kimi K3 language model, ensuring companies can manage intricate tasks involving deep reasoning skills.
Cohesive Ecosystem for AI
A scalable AI factory must rely on a consistent software ecosystem. Canonical’s Ubuntu provides a secure operating system layer, essential for smooth deployment and enterprise-level reliability. MiTAC partners with DDN to ensure that massive data streams are processed with ultra-low latency, preventing data shortages at the accelerator level. Furthermore, through collaboration with Rafay, MiTAC offers GPU-as-a-Service orchestration, which enhances resource allocation for distributed workloads, thus minimizing downtime and maximizing investment returns. The internally developed MiCoreView™ software consolidates this ecosystem, delivering detailed telemetry data and unified visibility throughout the data center.
Innovations in Infrastructure
MiTAC is also pushing boundaries by collaborating with AMD to develop the Helion Agentic AI rack system tailored for next-generation agentic architectures. This new solution is designed to handle an astounding 32,768 AMD Zen 6 cores within an open-source framework, enabling businesses to manage complex tasks more efficiently. At SC26, MiTAC debuted several server innovations for modern workload demands, including a high-density AI rack that integrates multiple servers to maximize GPU density by 50% compared to traditional deployments.
Conclusion
MiTAC Computing Technology Corporation offers a powerful suite of modular AI server platforms designed for customized enterprise deployments, showcasing its commitment to delivering energy-efficient, high-performance solutions. With a strong focus on quality and innovation, MiTAC is well-positioned to meet the evolving needs of businesses relying on cutting-edge AI and high-performance computing technologies. Attendees of the SC26 event are invited to visit stand 4615 to learn more about MiTAC's advanced offerings and the future of agentic AI integration. For further details, visit their official
website.