Werder Technology, a subsidiary of Quanta, has expanded its overall AI factory solution layout, extending from GPU servers to CPU computing, memory storage, network and infrastructure management. (File photo)
[Reporter Fang Weijie/Taipei Report] AI inference and agent-based artificial intelligence (Agent AI) applications continue to expand, driving the transformation of data center computing architecture. Quanta Technology (QCT), a subsidiary of Quanta (2382), has expanded its overall AI factory solution layout, extending from GPU servers to CPU computing, memory storage, network and infrastructure management, and integrating platforms such as NVIDIA, AMD and Intel to improve overall computing and energy efficiency.
Luo Weixian, director of products and solutions at Quanta and QCT, pointed out that AI computing requirements are changing with the development of reasoning, agent-based AI and long context memory. Complex reasoning requires more computing steps, and agent-based AI may perform tasks through multi-layer agents, further increasing the amount of token generation and making the AI infrastructure face higher computing and storage requirements.
Luo Weixian said that in the past, AI systems mainly focused on GPU computing capabilities, but as the inference workload increases, the importance of the CPU in task coordination has increased simultaneously. QCT therefore lays out high-density CPU cabinets so that CPU resources can be expanded independently, assisting GPUs in task allocation and reducing the situation where GPUs are idle waiting for work assignments.
In terms of memory architecture, Luo Wei first pointed out that AI computing generates a large number of Tokens and KV Cache (key-value cache). If all of them rely on expensive high-bandwidth memory (HBM) or local solid-state drives (SSD) with limited capacity, they will face resource allocation constraints. QCT uses low-latency, high-throughput storage solutions to help customers reuse the generated KV Cache, reduce repetitive operations, and release more GPU resources.
In addition to hardware architecture, Quanta has also deepened its cooperation with Huida's AI factory through QCT. Luo Weixian said that AI data centers must consider efficiency from design, construction, deployment to operation, and use Digital Twin technology to simulate power configuration, heat dissipation capacity and computing density in advance, and continuously monitor and adjust during the actual operation process to reduce deployment and operational risks.
Luo Weixian emphasized that the AI industry cannot only pursue increasing the number of GPUs, but must also think about how to use limited power to generate more tokens. Therefore, QCT simultaneously promotes efficiency optimization from both the hardware and infrastructure levels.
In terms of product layout, QCT not only provides AI factory software and hardware solutions with Huida as the core, but also integrates Intel and AMD platforms, covering CPU, GPU, storage and management systems, and uses self-integrated open source software solutions to provide functions such as rapid deployment, system management and real-time monitoring, allowing customers to choose the appropriate computing architecture according to different AI application needs.
