MiTAC Computing Unveils Scalable Full-Stack AI Infrastructure at SC26
In a significant showcase at the SC26 conference, MiTAC Computing Technology Corporation, a key subsidiary of MiTAC Holdings Corporation (TWSE: 3706), has introduced its cutting-edge AI infrastructure ecosystem. The exhibition, held at Booth 4615, carries the theme "Scaling Agentic AI. Built on Full-Stack Infrastructure," and illustrates how full-stack innovation can resolve critical bottlenecks in computational power, thermal regulation, and data handling that have hindered enterprise-level AI deployments.
Overcoming New Challenges in AI Deployment
As the demand for Agentic AI grows, evidenced by a considerable surge in inference workloads, the need for a robust infrastructure expands beyond mere GPU power. Addressing this, MiTAC collaborates with industry leaders such as AMD, Intel, and NVIDIA to deliver comprehensive frameworks that include high-density liquid cooling systems, disaggregated inference capabilities, and integrated orchestration solutions. This collaboration is pivotal in supporting the work of organizations looking to harness the full potential of AI technology.
Rick Hwang, President of MiTAC Computing Technology, emphasized that, "Scaling Agentic AI requires cohesive infrastructure, not fragmented systems." By providing a complete full-stack solution, MiTAC aims to transform the concept of sustainable AI factories from theory into practice.
Collaboration for Advanced AI Solutions
One of the highlights of MiTAC's presentation includes a joint project with Skymizer focused on developing a novel inference architecture. This architecture is designed to strategically allocate compute-intensive tasks to GPUs, while memory-intensive decode processes are handled by Skymizer's advanced HTX301 accelerator cards. This method allows enterprises to optimize each processing stage according to their specific workload needs, maintaining model inference capability on-site.
The collaboration involves two key server platforms:
1.
Prefill Stage: The MiTAC G8825Z5 (8U) server, equipped with AMD EPYC™ 9575F or 9755 processors, is designed for prefill processing. It efficiently transfers Key-Value (KV) cache data to the decode server via high-speed networking.
2.
Decode Stage: Utilizes the MiTAC G4520G6 (4U) server, which integrates Intel® Xeon® 6 processors alongside eight Skymizer HTX301-4S accelerator cards. This configuration boasts a remarkable 2 TB total memory capacity across its eight PCIe slots, enabling enterprises to run large models like the Kimi K3 language model effectively.
Enhancing Software Ecosystem
To complement its hardware innovations, MiTAC has structured a cohesive software ecosystem necessary for a scalable AI factory. Canonical Ubuntu serves as the reliable operating system foundation, offering assured deployment and enterprise-grade reliability. The integration of DDN Infinia further boosts data processing capabilities, ensuring minimal latency and preventing data bottlenecking, which often halts accelerator operations. Additionally, through a partnership with Rafay, MiTAC is capable of providing dynamic GPU-as-a-Service (GPUaaS) orchestration, optimizing GPU usage across various workloads.
The company's proprietary MiCoreView™ software acts as the linchpin, offering advanced telemetry and visibility across the entire data center, ensuring organizations can manage their resources effectively.
Showcasing Innovative Rack-Scale Solutions
MiTAC's advancements extend into rack-scale architecture developments as well. Building on the recent vision announced by AMD at the OCP Global Summit, MiTAC is developing the Helion Agentic AI Rack. This innovative solution is designed to support next-gen rack-scale architectures, scaling up to an astonishing 32,768 AMD Zen 6 CPU cores within an open-source framework. This capability is anticipated to enhance operational efficiency by intelligently routing complex workloads while optimizing agent performance per watt.
During SC26, MiTAC also highlights several innovative rack-scale solutions tailored for modern computational demands, including:
- - 52U High-Density AI Liquid-Cooled Rack: Incorporating 12 MiTAC G4826Z5 servers with 96 AMD Instinct™ MI355X GPUs, achieving impressive GPU density improvements.
- - Advanced Liquid Cooling Rack: Developed in partnership with ZutaCore, this proof of concept integrates rapid cooling technologies that manage thermal loads exceeding 4,000W efficiently.
- - High-Performance Storage Rack: Leveraging DDN Infinia’s intelligent AI data platform, this system ensures the seamless delivery of high-throughput data streams necessary for persistent AI operations.
Comprehensive Server Platforms for Custom Deployments
In addition to rack innovations, MiTAC is prepared to meet a diversity of enterprise needs with tailored server platforms:
- - AI Servers such as the G4827Z6 and G2610Z6, equipped with NVIDIA and AMD processors, are optimized for a range of inference workloads.
- - Enterprise Servers featuring the R1917GC and R1610Z6 are designed for commanding edge data center operations.
- - Multi-Node Servers like the M2810Z6 and C2611Z6 provide scalable solutions tailored for high-density and liquid-cooled data environments.
Attendees at SC26 are encouraged to visit MiTAC Computing at Booth 4615 from November 17-19 to learn more about their transformative full-stack AI infrastructure offerings.
For further details, you can explore MiTAC's offerings via their official website:
MiTAC Computing.