Aethlumis AI Infrastructure Solution

Background & Challenges

Modern enterprises and research institutions face exponential data growth and increasingly complex AI workloads.

• Conventional server systems are reaching their limits — struggling with:

• Insufficient GPU interconnect bandwidth, creating training bottlenecks

• Thermal inefficiency under sustained workloads

• Complex maintenance cycles with long downtimes

• Inflexible expansion paths that hinder scalability

Aethlumis addresses these barriers with an end-to-end intelligent computing solution that transforms traditional data centers into high-performance AI infrastructure.

Our Solution: Aethlumis TG990V3 Intelligent Compute Platform

The TG990V3 is Aethlumis’s next-generation AI flagship server, purpose-built for large-scale training, inference, and high-density data workloads.

It integrates cutting-edge hardware, modular architecture, and intelligent management, forming the core of our AI infrastructure stack.

Technical Highlights

• Compute Power: Dual 4th / 5th Gen Intel® Xeon® Scalable CPUs, TDP up to 350 W

• GPU Capability: Supports up to 8 OAM GPUs, fully interconnected under the OAI 2.0 standard

• Expansion Flexibility: Up to 14 × PCIe 5.0 slots + optional OCP 3.0 interface

• Storage Performance: Up to 20 × 2.5″ NVMe / SAS / SATA drives for high-throughput I/O

• Power Efficiency: Dual-plane design (6 × 54 V GPU zone + 2 × 12 V CPU zone) eliminates conversion loss

• Cooling System: 15 dual-rotor fans with zoned control, ensuring stable operation under 8 GPU full-load

• Smart Management: BMC AST2600 chip supporting IPMI 2.0, Redfish, and SNMP for full remote monitoring

This foundation enables a balanced topology architecture, supporting both High-Performance Dual-Uplink and Balanced Single-Uplink configurations to match your compute cluster requirements.

Solution Architecture Overview

Architecture Layers:

• Compute Layer — TG990V3 high-density nodes with 8 OAM GPUs

• Network Layer — 8 × 400 G interconnects ensuring ultra-low latency scale-out clusters

• Storage Layer — NVMe-based parallel storage for high-speed data access

• Management Layer — Unified Redfish/IPMI platform for orchestration, telemetry, and fault isolation

This modular, decoupled design allows independent upgrades, effortless maintenance, and horizontal scalability across racks or data centers.

Application Scenarios

AI Model Training

Designed for massive transformer-scale workloads, enabling large-parameter model training with minimal inter-GPU latency.

Supports GPUDirect RDMA and GDS for efficient data path between GPU and storage.

• Inference & Edge AI

Flexible GPU configuration allows inference acceleration for vision, NLP, or multimodal AI at scale.

Perfect for AI cloud services and on-prem edge deployments.

• Enterprise Compute Centers

Deploy TG990V3 as the backbone of your internal AI platform.

Unified management reduces O&M complexity and supports firmware orchestration, log collection, and smart diagnostics.

• Cloud & HPC Clusters

Seamless 400 G scale-out capability for large-scale compute fabrics — optimized for multi-tenant environments and hybrid AI clouds.

Integration Services

Aethlumis provides more than just hardware — we deliver complete AI infrastructure integration:

• Cluster design & deployment consulting

• Network topology optimization

• GPU resource scheduling & containerization (Kubernetes / Slurm)

• Thermal and power distribution design

• Remote management training & long-term support

Our engineering team works alongside your IT architects to ensure every watt, byte, and GPU cycle is fully optimized for your AI ambitions.

Partner Ecosystem

Aethlumis collaborates with leading ecosystem partners in compute, networking, and storage, including: Intel®, NVIDIA®, Broadcom®, Mellanox®, and Open Compute Project (OAI 2.0).

This ensures seamless compatibility and future-proof scalability for your investment.