MiTAC Computing Unveils Comprehensive AI Infrastructure Solutions at SC26

MiTAC Computing Technology's Vision for AI at SC26



At the recent SC26 conference held in Chicago, MiTAC Computing Technology Corporation showcased its latest advancements in AI infrastructure tailored for agent-based applications. As a subsidiary of MiTAC Holdings Corporation (TWSE: 3706), MiTAC has long been recognized as a leader in high-performance server solutions, and this event marked a significant leap forward in their mission to empower enterprises with scalable AI capabilities.

A Full-Stack Approach


Under the theme "Scaling Agent-Based AI: Built on a Full-Stack Infrastructure," MiTAC addressed the growing need for robust computational power in enterprise environments. Rick Hwang, President of MiTAC Computing Technology, emphasized that simply increasing GPU power is insufficient; a cohesive infrastructure is essential to overcome challenges related to computing capacity, thermal management, and data flow.

MiTAC's solution integrates strategic partnerships with industry giants like AMD, Canonical, DDN, Intel, NVIDIA, and Rafay to create scalable architectures specifically for AI factories. These innovations feature high-density liquid cooling systems, disaggregated inference, and comprehensive orchestration that together allow businesses to effectively handle the uptick in workloads driven by agent-based AI implementations.

Innovative AI Inference Architecture


A highlight from MiTAC's presentation was a groundbreaking collaboration with Skymizer, leading to an AI inference architecture that optimally assigns tasks between GPUs and memory-intensive accelerators. This balance will enable organizations to tailor setups according to specific workload requirements while allowing inference processing to remain within their on-premises installations.

Preload and Decoding Phases


The proposed configuration comprises two distinct server platforms:

1. Preload Stage:
- The MiTAC G8825Z5 (8U) system integrates AMD EPYC™ processors and AMD Instinct™ MI350X GPUs for effective preload processing and key-value cache transfer to decoding servers through high-speed networks.

2. Decoding Stage:
- The MiTAC G4520G6 (4U) system, equipped with Intel® Xeon® processors and Skymizer's HTX301 accelerator cards, can handle vast models like the Kimi K3 with its 2.8 trillion parameters. This setup fosters deep reasoning capabilities and autonomous workload handling in localized environments.

Comprehensive Software Ecosystem


Achieving scalable AI infrastructure also necessitates a coherent software ecosystem. MiTAC leverages Canonical’s Ubuntu as a foundational operating system, guaranteeing smooth deployment and enterprise-grade reliability. With DDN Infinia, MiTAC ensures minimal I/O latency while processing substantial data streams, effectively avoiding data starvation issues for its accelerators.

Collaboratively working with Rafay, MiTAC offers dynamic GPU resource orchestration as a service (GPUaaS). This solution maximizes return on investment by efficiently managing GPU resources across diverse distributed workloads.

MiTAC’s proprietary MiCoreView™ software enhances this ecosystem by providing real-time telemetry and unified visibility across data centers, reinforcing operational efficiency.

Rack-Scale Innovations


In alignment with industry advancements articulated by AMD's CEO Lisa Su at the OCP Global Summit, MiTAC is developing a next-gen Helion Agentic AI Rack system. This future-ready solution will support up to 32,768 cores of AMD Zen 6 CPU within an open code environment, enhancing the management of complex workloads while optimizing efficiency in agent performance per watt.

During SC26, MiTAC also revealed three specialized rack-scale innovations:

  • - High-Density AI Rack (52U): Integrating 12 MiTAC G4826Z5 servers with 96 AMD Instinct™ MI355X GPUs, this solution boasts a 50% higher GPU density compared to standard configurations.

  • - Advanced Liquid Cooling Rack: In collaboration with ZutaCore, a proof of concept was introduced that employs a bifasic liquid cooling system, achieving operational powers exceeding 4,000W with a thermal flux above 250W/cm².

  • - High-Performance Storage Rack: Built on the smart data platform for AI by DDN Infinia, this storage solution is designed for uninterrupted I/O operations, ensuring ultralow latency and high data throughput.

Expanding Server Platforms


In addition to rack architectures, MiTAC presented a range of fully integrated server platforms for custom enterprise deployments:

  • - AI Server G4827Z6: A modular 4U AI server optimized for NVIDIA MGX architecture, enabling diverse inference workloads.
  • - Compact GPU Server G2610Z6: This 2U server caters to edge environments using AMD EPYC™ processors, seamlessly supporting capabilities across network edges.
  • - High-Efficiency AI Platform TN85-B8261: Designed for specialized applications, this dual-socket platform integrates the AMD Instinct™ MI350P PCIe for effective AI acceleration.

MiTAC's commitment to innovation was evident at their booth (4615) during SC26 from November 17-19, inviting industry players to explore how their comprehensive solutions can redefine AI infrastructure in enterprises.

For more insights, visit MiTAC Computing Technology Corporation's website for extensive product details and latest innovations in server technology.

Topics Business 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.