In addition to hardware specification and rental pricing for the NVIDIA DGX H800 systems, many PolyU staff and researchers have requested operational details of the HK Cyberport platform to aid internal budgeting and research planning. The following information is provided for reference only and may not reflect the latest updates.
Since June 2025, a special subscription scheme is available starting from HKD 1,800 per month, open to all Hong Kong registered companies. Usage options range from 1/8 GPU to a full H800 card, with subscription duration from one (1) to three (3) months. If budget is permitted, users may consider this plan to explore the AISC’s feature and capabilities.
Download:
1. AISC Special Scheme Brochure
2. AIOP User Manual
3. Application Form (also available via Online Enquiry)
The following section outlines the technical requirements and limitations of the AISC. Students and researchers should carefully access whether these configurations and constraints align with their project needs and technical skill sets.
For More Information, please visit the Cyberport AISC Official Page for full details, or contact Louis CHAN (louischankk@cyberport.hk)
HK Cyberport AISC Platform and H800 Systems
The HK Cyberport AI Supercomputing Centre (AISC) is accessed through the Artificial Intelligence Operation Platform (AIOP), a secure web-based GUI for datasets uploads, resource management, and job monitoring. Access to the AIOP is strictly limited to pre-registered (whitelisted) IP addresses. Authentication requires a user ID, password, and mobile token (2FA). Session typically lasts several hours, and all computational operations are confined to the Cyberport AIOP VLAN to minimize attack surfaces.
Users interact with the platform through the AIOP portal, where they manage accounts, compute resources, storage, and AI applications. Workloads are launched either by provisioning Kubernetes containers or by submitting job via SLURM, the platform’s job schedular.
Please note that AIOP is intended primarily for intensive AI training and experimentation. Users are expected to fully develop and validate their code locally before submitting it to the AIOP. Production services are not permitted, as public internet access and TCP port exposure are disabled for security compliance.
AIOP operates as a secure platform for NVIDIA DGX H800 systems, making it well-suited for startups or research groups with large budgets and advanced technical capabilities. It offers integrated support for both Kubernetes and SLURM, supporting dynamic and efficient workload execution.
Internet Access Restrictions
Internet access on the AISC platform is strictly restricted to uphold high standards of security and data protection. By default, no internet connectivity is provided. Any request for access must be submitted on a per-domain basis with a valid justification. Approval is at the sole discretion of the AISC term.
Due to this no-internet policy, users cannot ping external sites (e.g., Google), install packages via Ubuntu's official online repositories, retrieve online datasets, or use cloud-based APIs during runtime. To support offline operations, the AISC team offers a shared mount disk containing pre-downloaded resources, including package repositories, Docker images, AI tools, and applications, allowing users to set up their computing environments without needing internet access.
If interest access is required (e.g., to reach specific HTTP domains), users must submit a formal request to the AISC team for individual domain whitelisting. Access to website categorized under social media, entertainments, file-sharing, personal email, and messaging platforms are strictly prohibited. Once approved, user may configure a proxy certificate to access the whitelisted domains. The total data transfer limit is 20GB for both uploads and downloads. For data exceeding this limit, user must utilize the AIOP portal or visit the onsite facility in person.
Additionally, the AIOP is an air-gapped system, meaning it is completely isolated from external networks. As a result, SSH and RDP access to provisioned resources is blocked, and no TCP ports are exposed. Users must interact with their deployed containers exclusively through the AIOP portal console. The platform is designed primarily for model training and is not intended for general development or debugging tasks.
Dateset Upload and Storage
All data and files must first be uploaded to the Redzone storage via the AIOP portal’s graphical interface, where they undergo mandatory scrutiny scanning. Each upload file is automatically scanned by AIOP's antivirus servers to ensure platform integrity. The standard upload limit is 50GB per file. For larger files, users must contact the AISC Support Team to arrange an onsite upload in person.
Once scanning is complete, users can manage their files by selecting, copying, or moving them into their project storage. Files stored in project storage can then be mounted and accessed by Kubernetes containers created by users.
Each rental DGX node in the HK Cyberport AIOP includes 5TB of NVMe-based high-performance storage (via DDN Lustre FS) and 10TB of SDD-based capacity storage (via Dell PowerScale) for datasets and workloads.
Workflow
- Upload to Redzone: Use the AIOP portal to upload datasets to Redzone (max. 50GB per file).
- Transfer to Project Storage: Move scanned files from Redzone to Tier1 (5TB/node) or Tier2 (10TB/node) project storage.
- Mount to Containers: Attach project directories from Tier1 or Tier 2 storage to Kubernetes containers for workload execution.
- Access via H800 GPU: Run AI computations on the mounted datasets.
Due to the air-gapped architecture, users cannot access their containers via SSH or SFTP.
All data uploads and downloads must be handled through the AIOP portal interface, which includes mandatory virus scanning and file staging workflows. Internet connectivity is disabled on the H800 systems, meaning users cannot retrieve online datasets or use cloud-based API during runtime. Alternatively, user may visit the AI User Experience Room located at Cyberport for direct data access, which require pre-registration and physical presence.
