Innovative Solutions for New Generation Liquid-Cooling AI Data Centers Based on NVIDIA GPUs

Author: William Ye, R&D Department, Aivres Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Aivres presents its end-to-end liquid-cooled AI data center solution portfolio based on NVIDIA B200 and B300 GPUs, spanning servers, racks, CDUs, secondary cooling loops, monitoring systems, and containerized modular deployment. The company offers two GPU server generations: the 5288 (B200-based, 5U with separate GPU and CPU nodes using AMD CPUs, 8 GPUs per system, 4+1 power redundancy) and the 5288 V3 (B300-based, supporting both Intel and AMD CPUs, with 1U OSFP board connecting 8 InfiniBand CX8 NICs, 10 NVMe SSDs).

For deployment, Aivres supports single-row POD designs of 16 racks with megawatt-class CDUs in 2N or N+1 redundancy. Containerized options include 20 ft containers (single-box with CDU, PDUs, monitoring, fire protection) and 40 ft containers (supporting four NVL72 racks). Four containerized boxes can be deployed within two weeks. Each rack holds 8 servers, with two manifolds and four PDUs mounted in the rear, providing 48 quick connectors per manifold. A notable European customer deployment consisted of a 20 MW data center with 6,144 GPUs in 96 racks delivering 13,000 PFLOPS, completed in 60 days with factory-preassembled secondary loop pipelines. The monitoring system provides 3D/2D visualization with leak detection requiring only RJ45 connections and no extra power, plus dynamic per-rack flow rate control based on thermal and power data.