# Ultimate Performance at National Scale

**Customer:** National Center for High-Performance Computing

NCHC designed a high-throughput, low-latency, scalable, and stable data and computing infrastructure centered on LLM training and inference needs, supporting AI research, development and services.

Introduction

## NCHC Eliminates Storage Bottlenecks to Build a High-Performance Computing Environment for LLM Training and Inference

Taiwan's National Center for High-Performance Computing (NCHC), under the National Institute of Applied Research (NIAR), is Taiwan's leading center for supercomputing, AI platforms, and high-speed network infrastructure. Since its founding in 1991, NCHC has driven the development of high-performance computing (HPC) in Taiwan, building and operating the Taiwania supercomputer series, which has repeatedly ranked on the global TOP500 list. NCHC also manages the Taiwan Academic and Research Network (TWAREN), delivering national-level education and research network services at speeds of up to 400 Gbps. NCHC's infrastructure supports universities, government research institutes, the semiconductor industry, and national AI programs across Taiwan, making it a critical platform for driving research innovation and industrial development.

As Taiwan's national AI strategy accelerates, NCHC has become a key pillar of sovereign AI development, supporting LLM training, climate science research, and quantum computing. As computing demand has grown, pressure on its storage infrastructure has risen quickly, and legacy systems have struggled to meet the performance and scalability requirements of today's largest AI and HPC workloads.

The Challenge

## When Storage Becomes the Performance Bottleneck

High-performance computing is only as fast as its slowest component. As AI workloads at NCHC grew in scale and complexity, its engineering team identified a critical challenge: storage I/O performance had begun to limit the full use of expensive GPU compute.

Te-Ming Chen, infrastructure engineer on NCHC's build and deployment team, has managed high-speed computing clusters and large-scale storage environments for years. But as AI applications expanded rapidly, the demands on storage systems changed fundamentally. "In high-performance computing environments, researchers are training and running inference with increasingly large models," Chen said. "The demand for high-throughput I/O continues to grow. Once storage systems can no longer keep up with compute speed, every research team waiting for resources is affected."

For LLM training, GPUs and storage systems are highly interdependent. Each training run repeats a cycle of data reads, GPU processing, and result writes, often thousands of times to complete a single model. When storage systems cannot sustain enough data throughput, GPUs sit idle waiting for data, wasting valuable compute and extending training time.

Inference workloads were equally affected by storage performance. Chih-Hao Chou, engineer on NCHC's LLM team, runs multiple production inference services. At startup, these services must load large model weights from storage into GPU memory. Under the previous storage environment, loading a single model took about a minute. That startup delay directly slowed service scale-up and reduced responsiveness for research users across Taiwan.

![Data center for NCHC](https://cdn.sanity.io/images/ult5g8gw/production/16614c8dd2ec9818988f0bf747c790f140cab988-1747x900.png)

The Approach

## Building a Data Platform for LLM Training and Inference

Following evaluation, NCHC deployed WEKA® NeuralMesh™, designing its storage and data services architecture around the actual requirements of its AI workloads: sustained high throughput for LLM training, fast model loading for inference, and the stability and maintainability expected of national-level computing infrastructure.

In an LLM environment, the data platform must support high-concurrency reads and writes across large numbers of small files and large model weights, while working in step with GPU clusters, high-speed networking, and existing job scheduling. The architecture needed to support parallel file access, horizontal scaling, low-latency model loading, data protection, and non-disruptive expansion, so the storage layer would not become a bottleneck in training or inference.

NeuralMesh also delivers real-time monitoring, capacity and performance visibility, anomaly alerts, and preventive maintenance, so the operations team can address issues before they affect users, reducing the risk of system interruption and improving overall service availability.

The Results

## Higher Research Efficiency, Greater GPU Value

### LLM Training Performance Improves

After deploying the new data platform, NCHC's team tracked improvements in LLM training workloads using metrics such as token throughput, GPU utilization, data read/write speed, and training cycle time. NeuralMesh acts as the high-speed data access layer in this architecture, helping storage and network systems deliver data more consistently, reducing GPU wait time, and making it easier for the overall training process to reach predictable performance.

For a shared computing platform, faster training means more than shorter completion times for a single job. It also improves overall resource turnover, letting more research teams get stable, predictable service quality from a limited pool of GPU resources.

### Inference Service Startup Gets Faster

For inference workloads, model weight load time, service startup time, and scaling speed are key measures of success. By building NeuralMesh into its data platform, NCHC can more effectively support fast reads and loading of large model weights, letting inference services start faster, scale elastically, and recover more quickly after maintenance, improving the experience for researchers and applications alike.

### Key Outcomes

- **Improved LLM training throughput**, raising GPU resource utilization
- **Reduced model loading times**, speeding inference service startup and scale-out
- **Shorter model development and validation cycles**, accelerating research output
- **Proactive monitoring and early-warning mechanisms **that reduce system interruption risk and improve overall service stability and operational efficiency

Looking Ahead

## Supporting Taiwan's National AI Mission

NCHC's infrastructure investments are more than technical upgrades. They represent a long-term commitment to Taiwan's research capacity and national competitiveness. Every gain in storage performance ripples across the shared computing platform used by hundreds of research institutions and thousands of researchers across Taiwan, further accelerating research innovation and output.

```json
{
  "_key": "00bd75de0ab5",
  "_type": "customerStoryMedia",
  "image": {
    "alt": "Rack space fo NCHC",
    "asset": {
      "_id": "image-89eff4f8f1a59182aded5ca2e901ababba3e6615-2700x2700-png",
      "metadata": {
        "dimensions": {
          "_type": "sanity.imageDimensions",
          "aspectRatio": 1,
          "height": 2700,
          "width": 2700
        },
        "lqip": "data:image/png;base64,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"
      },
      "url": "https://cdn.sanity.io/images/ult5g8gw/production/89eff4f8f1a59182aded5ca2e901ababba3e6615-2700x2700.png"
    },
    "crop": null,
    "hotspot": null
  },
  "layout": "image",
  "markDefs": null,
  "secondImage": null,
  "video": null
}
```

NCHC is also advancing plans for its future AI supercomputing system. Compared with the previous Taiwania 2 system, the new platform is expected to deliver more than eight times the AI compute performance, using NVIDIA HGX H200 and GB200 NVL72 systems connected by NVIDIA Quantum InfiniBand. It will support sover

eign AI, quantum computing, large-scale climate simulation, and the Taiwan AI RAP platform, providing generative AI development resources for startups, research institutions, and industry across Taiwan.

As GPU scale keeps growing and AI models keep getting larger, whether storage can scale in step with compute resources has become a defining question. NCHC's goal is to capture the full value of every GPU compute cycle and lower the technical barriers and wait times between a researcher's idea and its results.

The team at NCHC says the future supercomputer will drive breakthroughs in sovereign AI, quantum computing, and advanced scientific computing. It is designed to strengthen Taiwan's autonomy in critical technology fields, foster cross-domain collaboration, and improve Taiwan's competitiveness in global AI development.

By building NeuralMesh into its storage infrastructure, NCHC has created a scalable data platform for LLM training, inference, and the AI applications still to come. For a national computing center whose mission is to advance scientific research and technological innovation, efficient and stable storage infrastructure has become an essential foundation for that work.

[Download PDF](https://cdn.sanity.io/files/ult5g8gw/production/10ca7c0a9ca8a1c2f752e384baf5172d4f4ee720.pdf)

## Related

- [國家級規模的極致效能: NCHC 如何消除儲存 瓶頸，建置適用於 LLM 訓練與推論的 高效能運算環境](/resources/case-study/nchc-case-study-chinese-version)