MiTAC Computing Introduces Full-Stack Infrastructure for Agentic AI at SC26

MiTAC Computing Scales Agentic AI with Full-Stack Infrastructure at SC26



MiTAC Computing Technology Corporation, a subsidiary of MiTAC Holdings Corporation (TWSE: 3706), is making significant strides in the field of artificial intelligence (AI) by presenting its full-stack infrastructure for agentic AI at the SC26 event held in Chicago. In a world increasingly driven by AI workloads, MiTAC's innovative approach is designed to overcome traditional bottlenecks in computation, thermal efficiency, and data management, paving the way for smoother enterprise AI implementations.

During the event, MiTAC showcased its philosophy of "Scaling Agentic AI: Built on Full-Stack Infrastructure" at Booth 4615. This concept emphasizes a unified infrastructure instead of fragmented systems, revealing Rick Hwang, president of MiTAC Computing Technology. With the rise of agentic AI leading to overwhelming inference workloads, the company provides a comprehensive solution that integrates cutting-edge technologies from strategic alliances with industry leaders including AMD, Intel, NVIDIA, and Canonical.

Innovative AI Infrastructures



At the core of MiTAC’s presentation was the disaggregated AI inference engine developed in collaboration with Skymizer. This advanced architecture specifically delegates computationally intensive tasks to GPUs and resource-heavy decoding tasks to Skymizer's HTX301 accelerator cards. This configuration is tailored so enterprises can align their resources effectively across various phases of workload management, all while maintaining model inference on their own premises.

The proposed architecture includes two server platforms designed for distinct phases of operation:
1. Prefill Phase: Utilizing the MiTAC G8825Z5 server, equipped with AMD EPYC™ processors and AMD Instinct™ GPUs, for processing prefill tasks. This setup also enables efficient transfer of generated key-value caches to decoding servers.
2. Decode Phase: Using the MiTAC G4520G6 server, featuring Intel® Xeon® processors and equipped with HTX301 accelerator cards, designed for significant model deployments, such as the sophisticated Kimi K3 language model with 2.8 trillion parameters, which caters to complex agentic AI workloads requiring deep reasoning capabilities.

Comprehensive Software Ecosystem



For any scaled AI factory, a well-synchronized software ecosystem is crucial. MiTAC has integrated Canonical’s Ubuntu to ensure fundamental security coupled with seamless enterprise deployment. To manage vast data streams with minimal I/O latency, the DDN Infinia solution is employed, ensuring that accelerators receive adequate data supply.

Additionally, MiTAC collaborates with Rafay to provide dynamic GPU-as-a-Service (GPUaaS), efficiently allocating GPU resources to distributed workloads and thus maximizing return on investment while eliminating idle times. The binding element in this ecosystem is MiCoreView™, MiTAC’s proprietary software, delivering extensive telemetry data and overall visibility for the entire data center.

Showcasing Cutting-edge Innovations



MiTAC also introduced the Helion Agentic AI Rack in cooperation with AMD, marking a significant advancement in rack-level architecture for agentic systems. Targeted for next-generation architectures, the ORv3 44OU Helion solution is designed to scale to 32,768 CPU cores within an open-source stack while optimizing performance metrics.

Live demonstrations at SC26 are featuring three major rack innovations that cater to modern workloads:
  • - High-density Liquid-Cooled Rack: Featuring 12 servers with 96 AMD Instinct™ MI355X GPUs, this solution boasts a GPU density increase of 50% compared to traditional setups.
  • - Advanced Liquid Cooling Rack: A proof-of-concept showcasing elimination of traditional cooling methodologies, utilizing two-phase direct liquid cooling (DLC) techniques.
  • - High-Performance Memory Rack: Specifically catered towards providing uninterrupted data flow for agentic AI tasks without I/O limitations.

Furthermore, MiTAC is extending its platform offerings with a range of customizable server solutions tailored for businesses looking to implement optimal AI capabilities seamlessly. The introduction of various models, like the modular G4827Z6 and compact G2610Z6, highlights MiTAC's commitment to innovation and performance in the AI and HPC sectors.

Visit MiTAC at SC26



Industry professionals and exhibitors are encouraged to visit MiTAC Computing at Booth 4615 from November 17-19 at McCormick Place in Chicago. With a strong history dating back to the 1990s, MiTAC is uniquely positioned to offer energy-efficient server solutions that are essential for modern computing applications. Their dedication to quality, development, and integration assures partners the utmost reliability in advancing AI solutions.

For more information on MiTAC's offerings, please visit here.

Topics Consumer Technology)

【About Using Articles】

You can freely use the title and article content by linking to the page where the article is posted.
※ Images cannot be used.

【About Links】

Links are free to use.