Introduction The purpose of this RFI is to solicit input for a potential future ARPA-E research program focused on novel, potentially transformative technical opportunities, and approaches to increase the energy efficiency of data centers.
Online services and online data use have grown exponentially for years and is expected to grow even more so in the upcoming years due to growing trends such as virtual online meetings, artificial intelligence, machine learning, machine vision, and augmented reality1 . Recent indications also illustrate that we are on the precipice of accelerated machine-to-machine communication also known as Internet of Things (IOT) as cars, aircraft, industrial machines, and household devices will start gathering and communicating data for their operations2 . These virtual communications occur on an infrastructure supported by millions of servers that are hosted in data centers.
Data centers exist in a variety of sizes and operational models ranging from enterprise, co-location to hyperscale data centers. It is projected that current U.S. data centers consume in excess of 75 billion kWh electricity annually3 and that innovative technologies such as advances in semiconductors, virtualization and economization have contributed to a relatively modest increase in energy usage compared to the 550% growth in compute instances over the past decade4 . However, the semiconductor transistor scaling trend which consistently delivered operational energy savings during this time period, or “Moore’s law”, reached an inflection point in 2016, and it is therefore expected that data center energy usage will rise as future data communications demand is expected to grow exponentially5 .
The goal of the potential program is to support transformational technologies that can increase the energy efficiency of current and future data centers over as wide an application range as possible. Examples of potential areas of interest are disruptive innovations in efficient and reliable power supply, data processing, thermal management, server, rack, or building designs6 . These technologies could be developed as stand-alone component innovations or be considered as part of system-level approaches that span multiple component innovations as a system.
Traditionally, data center cooling has been one of the significant contributors to data center inefficiencies and its performance significantly climate dependent7,8 . Trends in the data center space such as higher power processors, increased power density per rack and the emergence of Edge data centers, which are placed closer to the customer to ensure low latency connections and high security for critical applications, are therefore expected to further exacerbate cooling and energy challenges9,10. Advanced cooling methods such as two-phase evaporation, and bio-inspired transpiration cooling11 have been explored for some time through implementations that include immersion cooling12 or specialized cold plate configurations13 but have only found limited commercial adoption. Cost and real or perceived reliability concerns have been mentioned as barriers for implementation of some of these new technologies in data centers.
Broad applicability can be improved by exploring technologies that deliver high performance at low cost and can operate efficiently, independently of the climate and/or water availability of the data center size or location. Technologies could explore the use of energy storage (thermal or electric) to take advantage of fluctuating demand/response and/or diurnal cycles.
Thermal management or cooling of a data center can arbitrarily be separated in three different optimization challenges and/or in a single system optimization challenge:
i. Move heat from current and future server chipsets to coolant efficiently
ii. Move heated coolant from server to the heat rejection system with high reliability
iii. Data center heat rejection systems that can reject heat efficiently or re-use the heat economically for auxiliary services that augment or benefit the data center.
An important objective of the data center operator is to have high availability, which is defined as the percentage of time the data center is available14. When the data center is considered as a system, failure mode and effect analysis can be used to assess the likelihood of occurrence, likelihood of detection, and severity of impact of potential failure modes of components and subsystems15. Generally, the lowest of the individual subsystem ratings determines the overall tier classification of the datacenter. ARPA-E would be interested to learn if new proposed transformational technologies would require health monitoring sensors and/or mechanisms for fail-safe operation, or advanced controls co-designed to ensure they have the potential to enable data center availability comparable or better than achieved with current technologies. In addition, it would be interesting to learn if prognostic health management and condition based maintenance could reduce server redundancy and further improve energy efficiency.
Another potential approach to increase total energy efficiency of data centers is to re-use the waste heat. Waste Heat Utilization for data centers has been an increasingly important area and increasingly encouraged and, in some cases, being explored to be mandated by local governments16. Although connecting a data center to a district heating system or green houses through co-location has been demonstrated, the mismatches in temperature, heat exchanger inefficiencies and operational offsets in energy demand and supply can be challenging.
As a potential program seeks heat rejection efficiency to be improved, heat could come out of the data center at higher temperatures than before (potentially slightly below typical chipset operating temperatures (70-90⁰C)). ARPA-E is interested to learn if energy at these temperatures (higher exergy/quality than current state-of-the-art) could potentially lead to new insights for energy re-use that would not need to rely on co-location of other heat demanding industries.
ARPA-E would be interested to learn if transformational technologies can reach additional performance if they were designed and/or co-optimized together. Such an increase could lead to an inflection point where the performance/cost ratio of new technologies would exceed baseline server technology at the data center level, and hence would be tech2market viable if simultaneously high reliability could be achieved.
The envisioned outcome of a potential program would deliver maturation of transformational technology concepts that display high potential performance and broad tech2market viability such that a leap in data center energy efficiency can be realized. Contact GMA if you have an interest in this RFI and the future solicitation.
Comments