Addressing the Data Center Power Challenge

Rapidly growing artificial intelligence (AI) and machine learning (ML) put a strain on data centers like never before. As a result, power consumption is rising significantly, especially across hyperscale facilities. Operators need to find ways to manage costs and conserve resources.

 

Advanced Energy’s Artesyn shelf and power modules.

 

Due to the proliferation of new advanced processors used to address AI and ML, the power-per-rack is rising dramatically. Data center operators now face the challenge of adding computing capacity while managing the increasing power usage.

There are widespread discussions on AI’s impact on society, privacy, and business. Its implications for power and the amount of energy required to feed the machine learning behind AI algorithms are less often mentioned. Adding to this increasing power demand is the plethora of cloud applications that require more power as they become more sophisticated.

These applications demand advanced processing in the form of microprocessors, GPUs, FPGAs, and ASICs, all of which are becoming more powerful with each generation. Microprocessors from the two main players, Intel and AMD, for example, have increased their power requirements by at least 50% in the short period between 2021 and 2024, reaching around 400 W per device.

Applications such as AI require increased use of GPUs. In the same period, GPUs from NVIDIA, specifically Grace Hopper, have increased their power consumption from 450 W to 1000 W, further driving up energy demand.

In the past, data center power-per-rack had been 6 to 15 kW. Today, 80% of rack power shipped is in the 20 to 40 kW per rack range. This is expected to more than double to 50 to 100 kW in the future. The power the already challenged grid consumes is only going in one direction rapidly.

 

Moving to 48 V Architectures

The data center challenge to provide more processing power in the same space and at lower cost seems impossible. While using the most efficient power conversion available will aid in reaching this goal, it can only go so far, and other enhancements must be considered.

Until recently, the primary power bus for distribution in racks was 12 V. Well-established, system components for 12 V are plentiful and, therefore, cost-effective. The challenge with 12 V is that as rack power increases in the relatively low-voltage bus, currents are correspondingly high, which drives conduction losses in the system.

Through initiatives such as the Open Rack V3 (ORv3) standard, driven by the Open Compute Project (OCP), more data centers are now moving to a 48 V bus to reduce conductive losses. 

The change to 48 V reduces the current by a factor of four for the same power. Consequently, cable distribution losses are reduced by a factor of 16 (I2R). However, additional benefits can also be realized – the reduced currents mean that cables and busbars can be smaller and thinner. This saves cost and space and allows power densities to increase. Power supply capacity can be reduced as less power is wasted. In addition, HVAC capacity and running costs are reduced due to less generated heat. 

In fact, there is an overall savings of around 1% in total cost of operation, just from changing to 48 V due to lower losses. Another 1.5% could likely be saved by power supply and system optimization.

 

48 V  Power Shelf Technology

Driven by the OCP ORv3 project, racks designed for 48 V and compatible power solutions are now available. Advanced Energy offers an ORv3-compliant rack power shelf in its Artesyn product line with 50 V/18 kW capability.

The 1OU open rack power shelf can be used with one or two power cords, allowing N+1 redundancy (15 kW) or N+N redundancy (9 kW) with dual cords. Input configurations can be 3 P Delta 4 W, 3 P Wye 5 W, and 3 x of 1 P.

A total of six 3 kW power modules can be fitted along with a hot-pluggable shelf controller for monitoring over Ethernet with optional battery backup.

Advanced Energy’s 3 kW open rack rectifier modules feature 97.8% efficiency and accept an AC input from 200 to 277 VAC. They are fully OCP-compliant and hot-pluggable to support redundancy within the rack.

The improvements and savings due to moving to 48 V or implementing more efficient power supplies can be significant at the scale data centers operate. 

For example, Advanced Energy’s ORv3-compliant power shelf delivers the industry’s highest efficiency (97.8%), at least 0.5% higher than the closest solution. It should be noted that this is substantially higher than many 80-PLUS Platinum and Titanium-graded power supplies. 

Even with a seemingly small 0.5% higher efficiency, continuous operation at 40% load would reduce the input power by 37.83 W (7.368 kW vs. 7.406 kW), which leads to an energy cost reduction of around $135 per shelf over five years, even for the most efficient data center, based on the U.S. Energy Information Administration’s 2021 reported national average of 7.18¢/kWh. Given the thousands of server racks in each data center, the seemingly small savings do add up in those big deployments.

 

Battery Backup and Peak Shaving

One significant challenge for data centers is to provide the capacity to deal with transient peaks. Additional capacity is needed to address occasional peaks to avoid overloads and circuit breakers tripping. However, over-budgeting power costs money and requires space. It also causes power supplies to run at lower than optimum capacity. 

Data centers store energy in batteries to cope with short-term outages in grid power. This is traditionally in large uninterruptible power supplies (UPSs) within data centers. However, in-rack battery backup units (BBUs) can be a more cost-effective and power-efficient way to provide DC power during an AC failure.

As the BBUs can be monitored and managed via an embedded controller, the energy stored within the batteries can be deployed to address short-term peak demand, allowing data center operators to limit the energy drawn from the grid. 

 

Using BBUs for peak shaving can reduce costs and space while improving power utilization.

 

This approach allows the creation of additional capacity within existing power envelopes, leading to improved power utilization when provisioning for peaks – reducing costs and space requirements. 

 

Hyperscale Data Center Challenge Summary

The power requirements for data centers are growing rapidly due to their importance in almost every aspect of modern life and the increased power requirements for emerging applications such as AI and ML.

Given the huge scale at which data centers operate, anything to enhance efficiency can be significant in terms of cost. Moving from 12 V to a 48 V primary bus voltage will have a considerable impact, which is why the transition is incorporated in OCP’s Orv3 specification.

Intelligent and managed use of energy stored in BBUs can ensure that power systems are efficiently provisioned, providing hold-up time while reducing cost and space and allowing them to operate more efficiently.

Addressing the hyperscale challenge is an area where Advanced Energy has contributed to the OCP ORv3 specification and developed best-in-class products to help engineers transition from 12 V to 48 V systems.

 

Images used courtesy of Advanced Energy