Resource Requirements

This page explains the infrastructure resource requirements and sizing guidelines needed when deploying the Aether Platform in an on-premises or private cloud environment. This is primarily intended for full on-premises deployments, joint research partners, or platform operators who need to perform independent hardware sizing and cost estimation.

[!NOTE] If you are using the SaaS version, all infrastructure management, sizing, and system operation functions are guaranteed by Aether and shared with other companies. Therefore, the considerations on this page are generally not required for SaaS users.

Aether Platform scales across three layers: Control Plane (Management), User Plane (IDE Execution), and AI Plane (Inference). It is strongly recommended to isolate the AI Plane, which consumes GPU resources, either physically or logically from other workloads.

Recommended Configuration & Sizing Requirements

Guidelines for choosing a configuration to run the Aether Platform effectively.

Configuration Summary (Includes 5x Safety Margin)

To guarantee a stable development experience and future scalability, the following specs are recommended as the “practical minimum.”

Feature	Minimum (PoC/Evaluation)	Recommended (Small Scale)
Target Usage	Individual evaluation, PoC	Small dev team (3-5 users)
Availability	None (Single Node)	HA (3-Node Cluster)
CPU (Total)	40 vCPU	120 vCPU (40 vCPU × 3)
RAM (Total)	80GB	240GB (80GB × 3)
Storage	1.2TB+ NVMe SSD	2.5TB+ NVMe SSD (Per Node)

[!IMPORTANT] The numbers above incorporate a 5x safety margin on top of the resources consumed by the OS, middleware, and base system. From a business continuity perspective, operation with specs below these values is not recommended.

[!WARNING] 250GB storage is the absolute minimum including OS, container images, and the base system. For actual development tasks (e.g., intensive Docker image builds or data accumulation), we strongly recommend allocating additional volume space.

1. Control Plane (Platform Overhead)

These are management functions shared by all tenants and users. Regardless of the number of users, the following fixed resources are required at a minimum.

Component	Minimum Specs	Recommended	Description
Kubernetes Master	3 Nodes (8 vCPU, 16GB RAM)	3 Nodes (16 vCPU, 32GB RAM)	Focus on high availability and responsive control plane.
Platform Services	3 Nodes (16 vCPU, 32GB RAM)	3 Nodes (32 vCPU, 64GB RAM)	Shared infrastructure (Identity, DB, etc.). Generous sizing recommended.
Gateway / NCS	3 Nodes (8 vCPU, 16GB RAM)	3 Nodes (16 vCPU, 32GB RAM)	Prepared for high traffic and security scanning loads.

Estimated Overhead: Maintaining the platform alone consumes approx. vCPU: 120–150 cores / RAM: 300GB+ to ensure consistent performance under any conditions.

Core Component Sizing (Resource “How Much”)

Guidelines for resource estimation to ensure stable operation of each core component.

1. User Plane: CloudIDE (Scales by Headcount)

The infrastructure for developer workspaces. “Concurrent user count” is the primary factor for determining resource requirements.

Type	Specs	Usage
Small	2 vCPU, 4GB RAM	Lightweight web development, scripting.
Standard	4 vCPU, 8GB RAM	Recommended. For most general development tasks.
Large	8 vCPU, 16GB RAM	Large-scale projects, compilation-heavy workloads.

Estimation Method: (Concurrent Users × Average Workspace Size) ÷ Overcommit Ratio (2-4)
Storage: Recommended 50GB+ NVMe SSD per user.

2. AI Plane: AI & Agent (Scales by Content/Usage)

AI agent performance depends on the “volume of knowledge” and “inference intensity.” Sizing is determined by how AI features are used rather than just headcount.

Feature	Scaling Factor	Recommended Specs
RAG (Search)	Knowledge volume (Docs)	Per 100k docs: 2 vCPU, 8GB RAM
Local LLM	Concurrent inferences	Per session: 1x NVIDIA T4/L4
AI Proxy	Request frequency	0.1 vCPU (When using external APIs)

Advice: If using RAG extensively, prioritize RAM for Vector DB. If running self-hosted models, adjust the number of GPUs accordingly.

3. NCS (Network Connectivity Service)

The zero-trust network overhead is designed to be extremely low.

Component	Resource Cost	Impact
NCS Gateway	8 vCPU, 16GB RAM	Per approx. 1Gbps throughput.
NCS Sidecar	0.5 vCPU, 512MB RAM	Proxy load for service-to-service mTLS.

Security: mTLS is applied to all communications. Optimization utilizing modern encryption instructions (like AES-NI) keeps application latency around 1ms.

Storage Requirements

System Disk (Minimum): For single-node or evaluation environments, at least 250GB is required for OS, container images, and the base system.
Block Storage (PVC): For each workspace’s /home directory. 50GB–100GB per user. High IOPS (e.g., NVMe over Fabrics) directly impacts development experience.
Object Storage (S3): For logs, backups, and Docker Registry. Low-cost storage is sufficient.
Vector DB: Requires fast read/write. Sizing should allow the index to fit in RAM.