Which liquid cooling solution performs best for GPU-intensive workloads: Direct-to-chip, immersion, or rear-door heat exchangers?
Which liquid cooling solution performs best for GPU-intensive workloads: Direct-to-chip, immersion, or rear-door heat exchangers?
Direct-to-chip (DTC) liquid cooling currently performs best for most GPU-intensive workloads, delivering high thermal efficiency for rack densities up to and beyond 132 kW. Immersion cooling suits the highest-density and specialised deployments, while rear-door heat exchangers (RDHx) provide the most practical path for retrofitting existing air-cooled facilities. The right choice depends on rack density, deployment scenario, and whether the facility is new or being upgraded.
Why Do GPU Workloads Need Liquid Cooling in the First Place?
GPU-intensive workloads generate heat that traditional air cooling cannot manage at scale. As accelerated computing clusters push rack densities beyond airflow limits, liquid cooling is the only viable means of maintaining stable, high-performance operation. Air cooling becomes insufficient once rack densities exceed 80–100 kW, which is now common in GPU-based AI training environments. At these power levels, airflow simply cannot extract heat quickly enough to prevent GPU throttling, the automatic performance reduction that occurs when a processor overheats. According to the International Energy Agency (IEA), global data centre electricity consumption is projected to more than double by 2030, driven largely by AI workloads. This growth makes cooling efficiency a commercial priority, not just an engineering one. Liquid transfers heat far more effectively than air, which is why the industry has moved decisively toward liquid-based cooling for high-compute environments.
What Is Direct-to-Chip (DTC) Liquid Cooling, and When Is It Best?
Direct-to-chip cooling circulates coolant through cold plates mounted directly onto CPUs, GPUs, and accelerators. It is currently the leading solution for both new-build and retrofit data centre deployments because it combines high thermal efficiency with practical integration. In our work across data centre deployments, direct-to-chip has become the default recommendation for GPU-dense environments, and the reason is straightforward. The cold plate sits directly on the heat source, capturing heat from the chip before it enters the air. This targeted approach handles extreme densities while allowing the rest of the facility to remain relatively conventional. Key characteristics of direct-to-chip cooling:
- Cold plates reject chip heat directly into the fluid distribution system
- Coolant Distribution Units (CDUs) manage flow rate, temperature, and pressure
- Manifolds distribute coolant evenly across servers for consistent cooling
- Supports single-phase and two-phase configurations as chip heat loads rise
Direct-to-chip is best suited for AI training clusters, high-performance computing (HPC), and any GPU deployment that requires predictable, high-density cooling without rebuilding the entire facility. Did You Know? Single-phase direct-to-chip cooling accounts for the majority of new liquid-cooled capacity today. Two-phase DTC adoption is expected to grow once chip thermal design power (TDP) exceeds the limits of single-phase designs, but for now, it remains largely at the pilot stage.
What Is Immersion Cooling, and Where Does It Win?
Immersion cooling submerges entire servers in a thermally conductive dielectric fluid, enabling uniform heat removal with minimal reliance on airflow. It delivers exceptional thermal performance and is used selectively where its specific benefits outweigh the operational trade-offs. Immersion cooling removes heat from every component simultaneously, not just the processor. This makes it highly effective for the densest deployments and for environments where consistent, whole-system cooling matters more than ease of maintenance. Immersion cooling removes heat from every component simultaneously, not just the processor. This makes it highly effective for the densest deployments and for environments where consistent, whole-system cooling matters more than ease of maintenance. However, immersion introduces several operational challenges that operators should consider:
- Staff may require additional training in the safe handling and maintenance of dielectric fluids, and there may be increased downtime during servicing because servers must be removed from the bath for upgrades or repairs.
- Compatibility of standard hardware with the immersion fluid also needs to be assessed.
- These factors mean that, despite its performance benefits, immersion cooling demands careful planning and may increase both maintenance complexity and operational risk compared to direct-to-chip solutions.
Immersion cooling is best suited for:
- Ultra-high-density deployments beyond the practical limits of direct-to-chip
- Environments prioritising maximum energy efficiency over maintenance simplicity
- Specialised HPC and research computing where whole-system cooling is advantageous
What Are Rear-Door Heat Exchangers (RDHx), and Why Retrofit With Them?
Rear-door heat exchangers replace the rear door of a server rack with a chilled-water-cooled radiator that captures server exhaust heat before it enters the room. They are the most practical option for adding liquid cooling to existing air-cooled facilities. RDHx units connect directly to the racks and complement liquid cooling while minimising impact on the white space, the operational floor area of the data centre. Because they do not require modifying the servers themselves, they are particularly well suited to phased upgrades and brownfield modernisation. For operators upgrading density incrementally rather than through a single large-scale replacement, RDHx offers a lower-disruption entry point into liquid cooling. It bridges the gap between fully air-cooled infrastructure and full direct-to-chip or immersion architectures. Rear-door heat exchangers are best suited for:
- Retrofitting existing air-cooled data centres without a full rebuild
- Facilities are transitioning gradually from air to liquid cooling
- Environments where minimising white-space disruption is a priority
Direct-to-Chip vs Immersion vs Rear-Door Heat Exchangers: A Comparison
The three technologies serve different priorities. This comparison summarises where each performs best across the factors that matter most for GPU-intensive deployments.
| Action | |||
| Factor | Direct-to-Chip (DTC) | Immersion Cooling | Rear-Door Heat Exchangers (RDHx) |
| Best density range | Up to 132 kW+ per rack | Highest densities, specialised | Up to ~72 kW per rack |
| Primary use case | New builds and retrofits | Ultra-high-density, specialised | Retrofitting air-cooled facilities |
| Retrofit friendliness | High | Low to moderate | Very high |
| Maintenance complexity | Moderate | Higher (fluid handling) | Low |
| White-space impact | Low to moderate | Moderate | Minimal |
| Maturity at scale | Highest, market-leading | Growing, targeted | Established for retrofit |
Which Liquid Cooling Solution Should You Choose for GPU Workloads?
The best solution depends on three questions:
- What is your target rack density?
- Is this a new build or a retrofit of an existing facility?
- How important is maintenance simplicity for your operations?
Matching these factors to the technology delivers the right outcome. For most GPU-intensive workloads in new or upgraded facilities, direct-to-chip is the strongest all-round choice; it handles extreme density while remaining practical to deploy and maintain. Immersion cooling earns its place at the highest densities and in specialised environments where its efficiency advantages justify the operational complexity. Rear-door heat exchangers are the pragmatic choice when the priority is adding liquid cooling to an existing air-cooled facility with minimal disruption.
In practice, many high-density data centres deploy a combination of direct-to-chip for GPU racks, rear-door heat exchangers for transitional zones, and air cooling for auxiliary spaces. Hybrid architectures are common because different zones within the same facility carry different density requirements.
How does Schneider Electric approach GPU Liquid Cooling?
We provide the full spectrum of liquid-cooling architectures rather than a single technology because GPU deployments rarely fit a single approach across an entire facility. This enables organisations to choose the best liquid cooling solutions for high compute and GPUs based on their infrastructure, workload density, and scalability requirements. Our liquid cooling portfolio is built around Coolant Distribution Units (CDUs) as the backbone between servers and heat rejection units. Liquid-to-liquid CDUs manage flow, temperature, and pressure control for the technology cooling system. In contrast, liquid-to-air CDUs allow liquid-cooled servers to operate within an air-based white space. Our Uniflair CDUs, cold plate accessories, and ChilledDoor rear-door heat exchangers cover direct-to-chip, hybrid, and retrofit scenarios, respectively. For AI-scale deployments, our reference designs, developed in collaboration with NVIDIA, support GPU rack densities up to 132 kW per rack, with validated guidance for both retrofit and purpose-built facilities. These designs provide operators with the technical specifications needed to accelerate high-density AI cluster deployments. Direct-to-chip liquid cooling can reduce total data centre energy use by 30% to 60%, improving Power Usage Effectiveness (PUE) and lowering operational emissions, a benefit that applies across our liquid cooling range.
Frequently Asked Questions
Air cooling generally becomes insufficient once rack densities exceed 80–100 kW, a common threshold in GPU-intensive AI training environments. Below approximately 30 kW per rack, well-designed air cooling with containment can remain effective. Between 30 kW and 80 kW, hybrid approaches such as rear-door heat exchangers bridge the gap. Above this range, direct-to-chip or immersion cooling becomes necessary for stable performance.
Yes. Most air-cooled facilities can be retrofitted without a full rebuild. Rear-door heat exchangers are the least disruptive option, as they replace rack doors without modifying servers. Direct-to-chip loops and CDUs can also be added to create hybrid architectures. Retrofit feasibility depends on the available floor space, the existing chilled-water capacity, and the target rack density.
Not universally. Immersion cooling delivers excellent efficiency at the highest densities but requires specialised fluid management and changes how servers are maintained. Direct-to-chip handles most GPU-intensive workloads with greater operational simplicity and easier integration, which is why it currently accounts for the majority of new liquid-cooled capacity. The best choice depends on density and operational priorities.
According to Schneider Electric's liquid-cooling data, direct-to-chip liquid cooling can reduce total data centre energy consumption by 30% to 60% when compared with traditional air-cooled systems. These efficiency gains contribute to better PUE, lower operational emissions, and stronger support for long-term sustainability and decarbonisation objectives.
Author Details:
Name - Anuj Kudesia
Email ID -Anuj.Kudesia@se.com
Company Name - Schneider Electric India
Linkedin Profile: https://www.linkedin.com/in/anujkudesia/
Author Bio: Anuj Kudesia, MD-SEPSL and Director of Business Development at Schneider Electric, leads strategy, product launches, and revenue operations for the Secure Power Division across the Greater India Zone. With over 20 years at Schneider Electric spanning enterprise data center sales, channel management, and strategic business development including P&L ownership for BFSI, IT/ITES, and Government verticals, he brings deep expertise in UPS systems, critical power infrastructure, edge computing, and data center modernisation for Indian enterprises.
Author Title: MD-SEPSL and Director of BD at Schneider Electric
ZZ_Paragraph Title
Need help?
Product Selector
Quickly and easily find the right products and accessories for your applications.
Get a Quote
Start your sales enquiry online and an expert will connect with you.
Where to buy?
Easily find the nearest Schneider Electric distributor in your location.
Help Centre
Find support resources for all your needs, in one place.