Thermosyphon Cooler Hybrid System for Water Savings in an Energy-Efficient HPC Data Center: Modeling and Installation (NREL/CP-2C00-66690, February 2017)

Overview

The paper frames heat rejection as a choice between dry sensible cooling, which needs no water but is bound by dry-bulb temperature and requires large plan area, and evaporative cooling, which is bound by lower wet-bulb temperature and removes far more heat per volume of air. For 28.3 L (1 ft3) of air and a 5.6 °C (10 °F) change, it quotes about 180 J dry against about 865 J in a tower. Because water and wastewater costs are rising and supply is less assured, and evaporative advantages shrink at cool ambient or low load, the authors argue the best option can be a hybrid. They describe a dry device in series upstream of the tower, so the system runs wet when hot and dry when not.

The thermosyphon cooler (TSC), made by Johnson Controls whose employees co-authored the paper, has a flooded shell-and-tube evaporator below an air-cooled condenser. Water boils the refrigerant around the evaporator tubes, vapor rises to the condenser, and liquid returns by gravity, so no pump or compressor circulates refrigerant. A controller sets condenser fan speed from ambient and process temperatures and the cost of energy and water. If load is lost in subfreezing weather, refrigerant flow stops and evaporator heaters keep the water from freezing, which lets the primary loop use plain water. The cooler's low pressure drop design lets water be cooled directly in the evaporator.

The host facility is NREL's HPC data center in the ESIF, designed for annualized PUE of 1.06 or lower and ERE of 0.9 or lower, with a 10 MW infrastructure ceiling, 2.5 MW installed and about 930 m2 of floor. Space conditions are 25.6 °C +/- 1.1 °C with dew points of 5.6-15 °C and RH up to 60%. Peregrine's 2,592 nodes (2.24 petaflops) take water up to 26.7 °C and return up to 40.6 °C, typically 37.8 °C. The operating load was about 0.91 MW, with 0.76 MW liquid-cooled. The TSC tied into the 350 mm energy recovery return pipe with only a one-day outage, using new 150 mm lines and valve, and sits on a platform meant for future towers.

Modelling used an hourly simulation across 8,760 hours of a typical year, with separate models for TSC, tower, exchangers and pumps, at $0.07/kWh and $1.61/m3 of burdened water. Building heat reuse peaks at 0.434 MW of 0.910 MW, leaving 0.496 MW to reject. The TSC alone could carry the full load below 9.4 °C dry bulb, about half the year. Tower make-up falls from slightly over 8,300 m3 to 3,700 m3 a year, saving 4,650 m3 or 56%, and utility costs drop 40%. The authors state these are initial model results, to be validated by tests reported in a follow-up paper.

Key points

1. Modelling projects annual cooling tower make-up of about 8,300 m3 (2.2 million gal) falling to 3,700 m3 (0.98 million gal) with the hybrid, a saving of 4,650 m3 (1.23 million gal) or 56%.
2. Modelled annual utility cost (energy plus water) falls by 40%, using $0.07/kWh energy and $1.61/m3 ($6.08/1,000 gal) burdened water.
3. The TSC can handle the entire load below 9.4 °C (49 °F) dry bulb, which occurs about 50% of the year at the site.
4. Per 28.3 L (1 ft3) of air, evaporative cooling removes about 865 J versus about 180 J dry for a 5.6 °C (10 °F) change, roughly 380% more.
5. The TSC circulates refrigerant by gravity with no pumps or compressors; fan speed is optimized for the economic balance of water saved and fan energy.
6. Loss of load in subfreezing weather stops refrigerant flow and energizes evaporator heaters, so the primary loop can run on plain water.
7. The ESIF HPC data center was designed for annualized PUE of 1.06 or lower and ERE of 0.9 or lower, with 10 MW capacity, 2.5 MW installed.
8. Space is held at 25.6 °C +/- 1.1 °C with dew points 5.6-15 °C and RH up to 60%; HPC return water is at or above 35 °C by design.
9. Integration needed only a one-day outage to cut the 350 mm return pipe and add a 150 mm branch and control valve.
10. Building heat reuse peaks at 0.434 MW of a 0.910 MW load, leaving 0.496 MW to reject to atmosphere.

Using it in design

• Place a dry stage upstream of the tower and size it against the share of hours where warm return water exceeds dry-bulb temperature by enough margin, as the 9.4 °C transition point does here.
• Run an hourly annual simulation with local weather and a stated water cost, because the case for hybrid rejection depends on burdened water price relative to energy price.
• Reserve tie-in points and valve space on the heat rejection loop so a dry stage can be added with a one-day outage, as the NREL pipework allowed.
• Verify freeze protection of dry devices on the primary loop, either through refrigerant-based isolation as in the TSC or glycol.
• Combine heat reuse first, dry rejection second, evaporative rejection last, and size towers for the hottest hours when the dry stage contributes little.
• Treat the numbers as predictions to compare against the measured results in the follow-up report.

Who should read it

Mechanical designers and sustainability leads evaluating hybrid dry and evaporative heat rejection for liquid-cooled or warm-water data centers.

Document contents

1. Abstract
2. Introduction
3. Traditional Choices for Rejecting Waste Heat to Atmosphere
4. Thermosyphon Cooler
5. Installation
6. HPC Data Center
7. HPC System Heat Rejection
8. Integration
9. Modeling
10. Conclusion
11. Acknowledgments

Publication details

Authors

Thomas Carter; Zan Liu; David Sickinger; Kevin Regimbal; David Martinez

Publisher

NREL

Year

2017

Document language

English

Source type

Government / public

Access

Citation only

Licence

Licence conditions not established — confirm the conditions with the publisher

Licence evidence

www.nrel.gov

Citation