Cloud computing has become a household term with growing significance and traction in the technology world. Simply stated, cloud refers to any service offered over the Internet. The concept behind cloud computing is to create a ubiquitous infrastructure to enable quick, scalable access to data and information. While most define and interpret the cloud to be a large public network, there are also private cloud services that offer secure proprietary networks with limited access and permissions.
Most consumers interact with the cloud through front-end access. The front end of the cloud includes software, applications, GUIs, and storage. To support the vast options of front-end user interfaces, the cloud requires a significant back-end infrastructure, including power supplies, servers, data storage and computers. With the ever-increasing demands of front-end cloud services, the back-end systems must also be scalable and capable of expansion.
According to ResearchAndMarkets.com, the global data center market is forecasting growth of 6.4% CAGR from $19.1 billion in 2020 to $26.1 billion by 2025. With the increasing growth in cloud computing demands comes the increasing demand on processing power. A recent study titled “Recalibrating global data center energy-use estimates” assessed that the worldwide power consumed by data centers in 2018 was 205 terawatt-hours, or 205,000,000,000,000 W-hr. Such significant demands in power consumption lend to the prioritization of efficiency and reliability.
Cloud power conversion
Most data center racks use an uninterrupted power supply (UPS) rated to 220 V with power ratings approaching 100 kW per rack. Considering most core processors have voltage ratings under 2 V, the high voltage levels need to be converted and distributed. Additionally, the higher power rating indicates significant amounts of current that need to be rerouted with utmost efficiency to minimize power losses and heat. Most server racks have a 48 V backplane power supply. It’s the primary supply into each server in the rack, also referred to as server blades.
Figure 1 The block diagram shows AC-mains to server backplane. Source: onsemi
It’s worth mentioning that 48 V has been the standard power supply in telecom and network infrastructure. The reason for selecting 48 V is that it’s generally considered to be the highest voltage that is not dangerous to humans. Typically, equipment requiring voltage levels over 48 V must be double insulated and have additional, stringent safety requirements. In addition, converting over 48 V requires ground isolation to protect equipment as well as humans operating them.
48 V vs. 12 V
There have been a lot of discussions and experimentation around the 48 V server power. The internal power supply in most computer and server platforms has been 12 V. It’s a legacy requirement that stemmed from older silicon technologies as well as hard disc drives for non-volatile storage, cooling fans, and other components in the compute platform. As CPU power consumption has dramatically increased with each successive generation of processors, the high CPU current load results in higher input 12 V current.
This higher current demand in turn requires thicker cables or bus bars to distribute 12 V and the higher 12 V current leads to larger distribution losses. Power losses also create heat which is the enemy of high-density computing as it leads to shorter device lifetimes and creates system vulnerabilities. One way to combat this power loss is to bring the 48 V rack supply into the server itself and introduce dedicated 48 V power converters.
Conduction power loss = (Load current2) x (Conduction path resistance)
The 48-V power supply can deliver the same power to a load with one quarter of the current; thus, reducing power loss in the conduction path by a factor of 16. This impressive improvement to system efficiency comes with some challenges. The 12-V power solutions have been optimized over many generations and are extremely efficient. On the other hand, the higher voltage power supply requires a larger step-down voltage to reach CPU core voltages, which can lead to a less efficient power conversion stage. Higher voltage silicon technologies are also required and tend to have higher resistance per unit-area for MOSFET architectures which will also increase system cost.
These system challenges have led to innovation and advanced architectures being implemented on a trial basis. One of the most promising new power conversion technologies is the switched tank capacitor (STC) converter shown in Figure 2. These converters exhibit extremely high efficiency and, in most cases, smaller circuit area. Depending on the designer and overall system architecture, both single and multi-stage conversion solutions have proven successful.
Figure 2 The switched tank capacitor (STC) converter comes in single and multi-stage designs. Source: onsemi
The specific intermediate voltage will vary by silicon vendor and is typically chosen based on specific technology requirements. The most efficient and widely chosen overall solutions have been 48 V to 12 V to 1 V to power the CPU core. This approach leverages both mature solutions and moderates the net step-down voltage to maximize total system efficiency (Figure 3).
Figure 3 Here is how the two-stage 48 V to 1 V converter works to power a CPU core. Source: onsemi
Smart power stages
High current DC-DC power converters are typically multi-phase topologies. Each phase typically comprises two MOSFETs in high-side and low-side half-bridge configuration and an inductor. The high-side and low-side MOSFETs are packaged together to increase power density and it’s commonly referred to as a power-stage. Multiple phases work together to deliver required output load and are controlled by an intelligent multi-phase controller.
Figure 4 A simplified view of of converter design that employs 16 phases to deliver 1 V CPU power supply. Source: onsemi
The switching of each phase must be staggered and carefully controlled to optimize load regulation, ripple, transient response and noise emissions, both radiated and conducted. The number of phases and the current in each one of those power stages are carefully tuned for a specific generation of CPU. The market has observed a rapid increase in the number of phases required as well as higher current density in each power stage. The most advanced multi-phase converters employ up to 16 phases with total delivered power easily exceeding 1,000 W (Figure 4).
A byproduct of the extreme power density required by advanced CPUs is the need for extremely high efficiency and tight load regulations. Advanced deep sub-micron silicon technologies employed in CPU/ASIC demand tight voltage tolerances in order to work properly. This drives the need for power stages to not only deliver low power losses, but also “smart” features including current, temperature, and fault reporting.
By reporting accurate phase current and temperature to the multi-phase controller, the overall power supply is able to deliver the required voltage regulation to the CPU. Advancements in MOSFET technology also play a key role in improving the efficiency in each generation of power stages. Figure 5 shows a more specific case of efficiency improvement. The designer needs to balance cost, peak efficiency, and maximum load efficiency for the best overall design.
Figure 5 The efficiency comparison of different power stages underscores the need for smart features. Source: onsemi
The cloud market segment will continue to evolve and expand as consumers expect more and more data at their fingertips. To keep up with these demands, the technology sectors supporting the cloud infrastructure must continue to innovate and anticipate the market needs. The entire cloud power tree including multi-phase controllers, smart power stages and POLs must be meticulously designed and manufactured to optimize efficiency and reliability to support this infrastructure.
Julie Tyler is strategic marketing manager at MCC Division of onsemi.