Continuing our series on advanced data center cooling solutions, this installment focuses on another innovative method to address the unprecedented thermal demands of artificial intelligence (AI) and high-performance computing (HPC) —direct-to-chip cooling using liquid cold plates and differential pressure control. After years of speculation that rack densities would one day reach 50 to 75 kW, the industry has far exceeded these expectations, with GPU rack systems now generating over 120 kW of heat. This trend is likely to continue, making the need for advanced cooling strategies that can manage extreme, rapidly increasing thermal loads a top priority for data center managers across the globe.
In this article, we explore how the Belimo Energy Valve paired with a differential pressure sensor is being used in world-class AI and HPC data centers to control direct-to-chip liquid cold plate cooling systems. Join us for a 20-minute webinar on Tuesday, March 18th, 2025, at 1:00 PM EST, where we will delve into the science behind this technology and its implications for the industry. Click here to reserve your spot.
Setting a New Standard
Some of the high-powered GPU server rack systems on the market come with guidelines that outline the best practices for cooling them. One major player in the industry recommends the Belimo Energy Valve paired with a differential pressure sensor as the optimal solution for managing cooling in their racks. Controlling differential pressure across a server rack is an entirely new approach to ensuring that flow rates remain constant through each cold plate, regardless of how many servers are installed in the rack at any given time.
Ensuring proper flow based on the number and type of servers installed to a rack can be challenging. When servers are removed for maintenance, a rack’s overall flow requirements can be disrupted, potentially causing overflow to the remaining servers if not properly compensated for. Replacing or upgrading servers can introduce even more variability, further complicating flow management.
Maintaining design flow across cold plates is critical for ensuring efficiency, preventing service outages, and preserving the longevity of both the cold plates and server components. Insufficient flow can result in inadequate cooling, leading to server damage and costly replacements, with high-performance models costing anywhere from hundreds of thousands to over a million dollars. Unplanned downtime caused by server freezes or unexpected shutdowns can amplify these losses, with the Uptime Institute’s 2022 survey reporting that 70% of outages cost over $100,000, and 25% exceed $1 million.
Excessive flow poses its own challenges, potentially eroding cold plate channels and compromising heat transfer efficiency. Higher flow rates also increase pump energy consumption, further escalating operational costs.
To address these challenges, starting with version 4.2, the Belimo Energy Valve introduces a new control option for maintaining a fixed pressure drop. This feature allows the Energy Valve to control direct-to-chip liquid cold plate cooling via differential pressure as opposed to traditional flow control. However, it requires a differential pressure sensor (not included with the Energy Valve), which must be wired to the EV.
Flow Control vs. Differential Pressure Control
Figure 1: Flow Control with the Energy Valve
Using traditional flow control, the Energy Valve is configured to control the total flow to the rack. To determine the total flow required to cool the entire rack, the flow demands of each individual cold plate must be summed. This technique ensures stable and consistent cooling performance, unless one of the servers needs to be removed. In this scenario, the flow would become too high for the remaining servers.
Figure 2: Differential Pressure Control with the Energy Valve and differential pressure sensor
On the other hand, maintaining constant differential pressure across the rack with the Energy Valve and a differential pressure sensor ensures consistent flow to each cold plate, even as servers are brought on and offline. When a server is removed from the rack, the overall differential pressure will increase. The Energy Valve will sense this and reduce the flow rate until the programmed pressure drop is reached. Since differential pressure and flow are directly proportional with a fixed resistance, once the prescribed pressure drop is reached, the flow will automatically adjust to the correct level for the remaining servers.
Conclusion
As data centers face the growing challenge of managing extreme heat densities, differential pressure control with the Belimo Energy Valve has become the industry standard in liquid cooling for high-performance GPU server racks. This technology ensures consistent flow across cold plates, even as server configurations change, optimizing cooling performance, reducing operational risks, and improving overall system efficiency.
Join us for a 20-minute webinar on Tuesday, March 18th, 2025, at 1:00 PM EST, where we will explore the science behind differential pressure control and how it’s used in AI/HPC data center cooling. Click here to reserve your spot and discover how this solution can future-proof your operations.