GIGABYTE Direct Liquid Cooling Solution

Productivity at Scale for the AI Factory Era:
Servers × Liquid Cooling × G-REX

Generative AI, large language models (LLMs), and high-performance computing (HPC) are reshaping modern data centers into AI factories. As rack power density reaches 100 kW and beyond, scaling compute is no longer just a performance challenge, it is an energy, cooling, and operational challenge.

According to a 2024 report from the U.S. Department of Energy (DOE) and Lawrence Berkeley National Laboratory (LBNL), data center power demand will continue to grow rapidly, placing sustained pressure on electrical and cooling infrastructure. To scale AI sustainably, cooling must be designed as a first-class system, not an afterthought.

Direct Liquid Cooling: Designed In, Not Bolted On

Unlike retrofit liquid cooling kits for air-cooled servers, GIGABYTE Direct Liquid Cooling (DLC) servers are engineered for liquid cooling from day one. High-heat components are coupled directly to the liquid loop through cold plates, enabling the majority of system heat to be removed at the source. Compared to air cooling, which relies on high airflow and energy-intensive room cooling, DLC shortens the thermal path, improves efficiency, and significantly reduces cooling overhead.

>92% Heat Capture Ratio: Built for High-Density AI and GPU Clusters

Heat Capture Ratio (HCR) is a key indicator of liquid cooling effectiveness. GIGABYTE DLC servers remove over 92% of total system heat directly through liquid cooling, leaving minimal residual heat to be handled by air.

Operational benefits include:

Reduced or eliminated internal server fans
Lower data center cooling and chiller demand
Improved PUE and overall energy efficiency
Increased stability for sustained AI training and high-power GPU workloads

Independent studies and field evaluations confirm that cold-plate DLC is a mature technology. DLC has demonstrated stable operation under 140 kW rack-level IT loads, making it suitable for large-scale AI and HPC deployments.

GIGAPOD: Modular, Scalable AI Infrastructure

In production environments, the real challenge is not deploying a single high-performance GPU server, it is operating hundreds of racks consistently and reliably. GIGAPOD extends GIGABYTE’s DLC technology into a fully integrated, liquid-cooled compute rack solution. Servers, networking, power, and cooling are pre-integrated and validated at the factory, enabling predictable performance and faster time to deployment.

Name	Positioning	Application	Details	Features
GIGAPOD AI	High‑performance GPU architecture built for AI training and inference	Large‑scale AI training, generative AI inference, LLM training clusters	HGX / OAM/ PCIe accelerator cards	Extreme AI performance, optimized thermal efficiency, turnkey deployment
GIGAPOD HPC	Scalable compute architecture optimized for HPC and cloud workloads	Scientific simulation, engineering analysis, cloud computing, national labs / university HPC	CPU‑based clusters with multi‑node configurations	High stability and efficiency, modular deployment, support for MRDIMM and high‑speed networking

Key Advantages:

Shared cooling architecture across multiple racks
Linear scalability aligned with compute density
Centralized management for AI factory operations
Support for rapid deployment (L11 / L12 delivery models)

GIGAPOD is designed as a standardized building block for enterprise AI factories, cloud service providers, and research institutions, reducing integration complexity and accelerating expansion.

Warm Water Cooling up to 45°C, Reducing Chiller Dependency and Energy Cost

Energy efficiency at scale often comes from minimizing chiller usage. Industry studies show liquid cooling systems operating at near-ambient or elevated supply temperatures can reduce chiller runtime and extend free cooling availability.

GIGAPOD is engineered to support stable operation with inlet water temperatures up to 45°C, allowing compute capacity to scale without proportional increases in energy usage.

This aligns with modern data center best practices: lower water temperature does not equal higher efficiency, system-level optimization does.

G-REX: Unified Operations for OT and IT Infrastructure

Liquid-cooled data centers introduce a growing number of operational technology (OT) systems, including CDUs, rear-door heat exchangers (RDHx), and environmental controls. Without integration, these systems create operational complexity. G-REX is GIGABYTE’s middleware platform and a core component of GIGABYTE POD Manager (GPM). It consolidates cooling and power infrastructure data and presents it through IT-friendly, cluster-level interfaces.

Cluster-Level Visibility and Control

Real-time monitoring across racks and subsystems
Unified view of cooling, power, and environmental data
Seamless integration with higher-level cluster management platforms

Open Standards and APIs

G-REX supports Modbus, Redfish, and REST APIs, enabling integration with in-rack, in-row, sidecar CDUs, and rear-door heat exchangers. Customers can adopt liquid cooling incrementally without locking into a single vendor ecosystem.

Built for Enterprise Productivity:
A Complete Path from Deployment to Operations

In the AI era, infrastructure is a strategic productivity asset. Enterprises require solutions that are scalable, energy efficient, and operationally predictable. By combining DLC servers, GIGAPOD rack-scale systems, and G-REX cluster operations, GIGABYTE delivers a fully integrated AI infrastructure platform, from deployment to long-term operations. The result is lower total cost of ownership, faster time to value, and sustainable AI scalability, the foundation enterprises need to operate AI factories with confidence.

Liquid-Cooled GIGAPOD Configurations

GIGAPOD AIGIGAPOD HPC

GPU Solutions	Supported Systems	Systems per Rack	Rack Height	Racks per Scalable Unit (Compute + Management)	Power Consumption per Rack	PDU per Rack	CDU
NVIDIA GB300 NVL72	DLA2-CB3	18	48U	16+8	140kW	6 x 63A	In-Rack
NVIDIA GB300 NVL72	DLA2-CB3	18	48U	16+8	140kW	6 x 63A	In-Row
NVIDIA GB200 NVL72	DLA2-CB0	18	48U	16+10	130kW	6 x 63A	In-Rack
NVIDIA GB200 NVL72	DLA2-CB0	18	48U	16+10	130kW	6 x 63A	In-Row
NVIDIA HGX B300	G4L4-SD3-LAX7 G4L4-ZD3-LAX7	8	42U	9+4	109kW	6 x 63A 2 x 125A	In-Rack
	G4L4-SD3-LAX7 G4L4-ZD3-LAX7		42U	9+3	109kW	6 x 63A 2 x 125A	In-Row
	TO46-SD3-LA07		44OU	9+4	90kW	6 x 63A	In-Rack
	TO46-SD3-LA07		44OU	9+3	90kW	6 x 63A	In-Row
AMD Instinct MI355X	G4L3-ZX1-LAT4	8	48U	4+1	120kW	4 x 100A	In-Rack
AMD Instinct MI355X	G4L3-ZX1-LAT4	8	48U	4+1	120kW	4 x 100A	In-Row
NVIDIA HGX B200	G4L3-ZD1-LAX5 G4L3-SD1-LAX5 G4L4-SD1-LAX5 G4L4-AD1-LAX5	8	42U	4+1	103kW	4 x 100A	In-Rack
NVIDIA HGX B200	G4L3-ZD1-LAX5 G4L3-SD1-LAX5 G4L4-SD1-LAX5 G4L4-AD1-LAX5	8	42U	4+1	103kW	4 x 100A	In-Row
AMD Instinct MI325X	G4L3-ZX1-LAX2	8	48U	4+1	110kW	4 x 100A	In-Rack
AMD Instinct MI325X	G4L3-ZX1-LAX2	8	48U	4+1	110kW	4 x 100A	In-Row
NVIDIA HGX H200	G593-SD1-LAX3 G593-ZD1-LAX3	8	48U	4+1	85kW	4 x 63A	In-Rack
	G593-SD1-LAX3 G593-ZD1-LAX3		48U		85kW		In-Row
	G4L3-SD1-LAX3 G4L3-ZD1-LAX3		42U		80kW		In-Rack
	G4L3-SD1-LAX3 G4L3-ZD1-LAX3		42U		80kW		In-Row
AMD Instinct MI300X	G593-SX1-LAX1 G593-ZX1-LAX1	8	48U	4+1	85kW	4 x 63A	In-Rack
	G593-SX1-LAX1 G593-ZX1-LAX1				85kW	4 x 63A	In-Row
	G4L3-ZX1-LAX1				90kW	4 x 100A	In-Rack
	G4L3-ZX1-LAX1				90kW	4 x 100A	In-Row

Rack Height / Node Count	Supported Systems	Systems per Rack	Power Consumption per Rack	PDU per Rack	CDU
42U / 60 Nodes	H174-A80-LAS1	30	105kW	4 x 100A	In-Rack
In-Row
H274-A81-LAZ1 H274-S61-LAW1	15	In-Rack
In-Row
42U / 50 Nodes	B683-Z80-LAS1 B684-A80-LAS1	5	90kW	4 x 100A	In-Rack
In-Row