⚡ electrical engineering

IT-power feedforward halves pump energy in liquid-cooled data centres

Replacing temperature-based feedback with a direct feedforward control strategy using IT power as the input variable reduces pump energy consumption by over 56% and peak junction temperatures in liquid-cooled data centers, demonstrating that the observed variable is more critical to efficiency than algorithmic complexity.

Original authors: Yang Liu

Published 2026-10-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Liu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside the massive server rooms that power our digital world, a quiet revolution is taking place in how we keep the machines from overheating. As artificial intelligence grows more powerful, the computer racks that run these systems are becoming incredibly dense, packing so much computing power into a single cabinet that traditional air conditioning can no longer keep up. To solve this, engineers have turned to liquid cooling, where a special fluid is pumped directly over the hot computer chips to carry the heat away. However, this solution brings its own problem: the pumps that move this fluid are now the biggest energy consumers inside the cooling system itself. The energy these pumps use does not grow in a straight line; instead, it grows explosively. If you double the speed of the water flow, the energy the pump needs to push that water increases eight times. This means that finding the absolute minimum amount of flow needed to keep the chips safe is the single most important way to save electricity. For decades, the standard way to control these pumps has been to watch the temperature of the coolant. The pump waits until the water gets warm, and only then does it speed up to cool things down. This approach is reactive, meaning the system is always chasing the heat after it has already appeared.

A new study by independent researcher Yang Liu challenges this long-held method by proposing a much simpler, faster approach. Instead of waiting to see the temperature rise, the researcher suggests the pump should simply watch the electricity being used by the computer chips. Because the heat a chip generates is directly tied to how much power it consumes, the system can calculate exactly how much water flow is needed before the heat even arrives. In a detailed computer simulation of a high-power server rack, this method of looking ahead at the power usage—rather than reacting to the temperature—cut the energy used by the pumps by more than half. Specifically, the new method reduced the pump's energy consumption by 56.2 percent compared to a highly tuned traditional controller that was adjusted over ninety different settings to perform as well as possible. The simulation showed that this new approach not only saved massive amounts of electricity but also kept the computer chips cooler, lowering their peak temperature by more than five degrees Celsius.

The research demonstrates that the key to efficiency lies in what the controller observes, not in how complex the computer algorithm is. The proposed system uses a direct calculation based on the power reading, requiring no complex forecasting models or learning from past data. It simply reads the power, calculates the necessary water flow, and adjusts the pump speed immediately. In contrast, the traditional temperature-based controller has to wait for the water to heat up, then increase the flow, and then wait for the temperature to drop again. This delay creates a cycle where the pump often moves too much water when it is not needed, wasting energy, or too little water during sudden spikes in activity, risking overheating. The study found that the new method kept the water flow between 26 and 86 liters per minute, whereas the traditional method swung wildly between 14 and 117 liters per minute, often overshooting the mark.

The researchers also tested a common idea in engineering: that looking further into the future to predict a load increase would help the system prepare. They simulated a scenario where the pump would start speeding up minutes before a computer task began, hoping to be ready for the heat. The results were surprising. Looking ahead by up to five minutes actually increased the energy consumption by 45.5 percent without providing any cooling benefit. This happened because the water flow needed is an immediate response to the power being used; adding extra flow before the power rises serves no purpose since the temperature is already safe. The study concluded that the only time looking ahead would be useful is if the pump itself were incredibly slow to respond, which was not the case in this setup. The researchers also explored whether moving water from one server rack to another could save energy, but found that each rack has its own strict minimum flow requirement based on its own heat load. Trying to take water from one rack to give to another would only break the safety limits of the donor rack, proving that energy savings come from meeting each rack's needs precisely, not by shuffling resources around.

This work, while currently a computer simulation rather than a physical experiment, offers a clear path forward for the design of future data centers. The findings suggest that the most effective way to reduce energy use is to stop treating the cooling system as a reaction to temperature and start treating it as a direct response to power consumption. By using a simple, five-step rule that links power usage directly to water flow, data centers could drastically cut their electricity bills and reduce their environmental footprint. The study emphasizes that the solution does not require smarter, more complex software, but rather a smarter choice of what to measure. In a world where data centers are projected to double their electricity demand in the coming years, shifting the focus from the result of the heat to the cause of the heat provides a straightforward, powerful tool for efficiency.

Technical Summary: IT-Power Feedforward in Liquid-Cooled Data Centres

Problem Statement
In modern liquid-cooled data centres, particularly those supporting high-density AI clusters (50–120 kW per rack), the pump circulating coolant has become the largest single energy consumer within the cooling loop. Pump power consumption scales with the cube of the flow rate ( $P_p \propto Q^3$ ). Consequently, the choice of the controlled variable fundamentally sets the achievable energy floor. Current industry practice relies on temperature feedback control, where pump speed is adjusted based on the coolant temperature or chip junction temperature. This approach is reactive: the controller waits for a thermal excursion (the effect) to occur before increasing flow, despite the fact that the heat load (IT power, the cause) is directly measurable in real-time at the power distribution unit (PDU) or server baseboard. This paper investigates whether replacing temperature feedback with a direct feedforward control based on IT power can breach the energy floor established by reactive temperature control.

Methodology
The study utilizes a one-dimensional, lumped-parameter numerical model of a 90 kW rack under a specific load profile (a step from 30% to 100% load, followed by a drop to 40%). The model is governed by three physical relations: energy conservation, the quadratic network characteristic ( $H = SQ^2$ ), and the resulting pump power law ( $P_p \propto Q^3$ ).

Three hard constraints define the feasible operating space:

Junction Temperature ( $T_{chip}$ ): Must not exceed 85 °C.
Supply-Return Temperature Difference ( $\Delta T$ ): Must not exceed 15 °C.
Actuator Slew Rate: The flow rate change is limited to the pump's full range over 60 seconds.

The author compares three control strategies:

Temperature Feedback (Baseline): A proportional controller ( $q_{cmd} = Q + k_p(T_{chip} - T_{sp})$ ) tuned over a grid of 90 parameter sets ( $T_{sp}$ and $k_p$ ). The lowest-energy configuration that satisfies all constraints is selected as the benchmark.
IT-Power Feedforward: A closed-form control law that calculates the required flow instantaneously based on the measured IT power ( $P_H$ ). The flow demand is derived algebraically from the inverted constraints: $Q \geq \max(Q_{S1}, Q_{S2})$ , where $Q_{S1}$ ensures $T_{chip} \leq 85$ °C and $Q_{S2}$ ensures $\Delta T \leq 15$ °C. This controller contains no feedback path, no PID loop, and no forecasting.
Look-Ahead Strategies: Variations of the feedforward controller that anticipate future load increases (0 to 600 seconds ahead) to test if pre-emptive flow elevation yields energy savings.

Additionally, a multi-rack scenario (80 kW, 50 kW, and 30 kW racks sharing a common header) is simulated to test the viability of inter-rack flow reallocation for energy savings.

Key Results

Energy Reduction: The IT-power feedforward controller reduced pump work by 56.2% compared to the optimally tuned temperature feedback controller (0.0765 kWh vs. 0.1747 kWh) over the 1800-second simulation.
Thermal Performance: The feedforward approach also lowered the peak junction temperature by 5.14 °C (78.50 °C vs. 83.64 °C).
Flow Dynamics: The feedback controller exhibited both under-supply during transients and over-supply in steady state. The feedforward controller maintained flow strictly within the necessary bounds (26–86 L min⁻¹), avoiding the excessive flow (up to 117 L min⁻¹) seen in the feedback loop.
Look-Ahead Penalty: Implementing a look-ahead strategy (raising flow before a load step) resulted in a negative return. A 300-second look-ahead increased pump work by 45.5% without reducing peak temperature. This is because flow demand is an instantaneous function of power; adding flow before the power rises serves no constraint, and the actuator speed (60 s ramp) is sufficient to handle the step without pre-emption.
Inter-Rack Reallocation: Attempting to save energy by reducing flow in one rack and increasing it in another failed. While total pump power decreased slightly in simulation, the donor rack immediately violated its 15 °C temperature-difference constraint. The study concludes that lower bounds are set by individual rack loads and are not transferable resources.

Significance and Claims
The paper argues that the primary source of energy inefficiency in liquid cooling is not algorithmic sophistication but the choice of the observed variable.

Cause vs. Effect: Temperature is the result of the thermal balance; controlling it requires waiting for an error to accumulate. IT power is the cause; its relationship to required flow is purely algebraic. By observing the cause, the controller prevents the deviation from arising rather than correcting it faster.
Simplicity: The proposed solution requires only five statements of code, no PID tuning, no machine learning, and no predictive models. The gain is derived from the physics of the system and the availability of real-time power data, not from complex control theory.
Structural Limits: The study demonstrates that inter-rack coordination cannot save energy if every branch must independently satisfy its own thermal constraints. The minimum energy state is the vector of individual lower bounds.
Operational Insight: The binding constraint in the tested parameters was the 15 °C supply-return temperature difference, not the 85 °C junction limit. This suggests that $\Delta T$ specification is a first-order operating parameter for energy efficiency, rather than a mere convention.

The author explicitly states that these findings are based on numerical simulations using a lumped-parameter model and do not include hardware interfacing, redundancy, or site-specific variables. However, the results indicate a clear magnitude and trend: replacing temperature feedback with IT-power feedforward offers a substantial, immediate reduction in auxiliary cooling energy.

Technical Summary: IT-Power Feedforward in Liquid-Cooled Data Centres

More like this