Organizations like Open Compute Project (OCP) and Open19 have created open standards for alternatives to conventional server designs. These architectures redefine how power is converted and distributed within an IT rack. Centralized rack-level power supply units (PSU) replace internal server power supplies. 12V has been the standard output for power supplies, but recently we’re seeing 48V, with claims of much higher overall efficiency.
Why 48V? It comes down to the quest for improved compute performance, which requires increasing the power to the chip, and at 12V, this isn’t practical. I’ll explain why in another blog.
Efficiency: A Compare and Contrast Story
We thought it was important to research this topic to understand the expected efficiencies from the PSU to the server point of load (the components within the servers). We created a model to illustrate the differences between the architectures, show the main drivers of the electrical losses, and see how varying parameters like redundancy of PSUs, oversizing of PSUs, and component efficiencies can impact the resulting overall efficiency.
Try the Rack Power Architecture Efficiency Calculator in our Tradeoff Tool Library.
The calculator allows you to analyze the power path efficiency of the different architectures, starting from the input power of the rack (PSU input), down to the voltage regulator modules (VRM) within the servers that deliver power to the individual server components. Specifically, we looked at comparing conventional servers (PSUs within each server) and centralized rack-level PSUs, both 12V & 48V DC approaches.
With the default settings in our tool, the efficiency of the 12V rack level PSU architecture is 7 percentage points better than conventional servers (or 33% reduction in losses), and 48V is just over 1 percentage point better than 12V rack-level (7% reduction in losses).
So where does the efficiency improvement come from?
- Redundancy – Centralizing PSUs to the rack allows for N+1 redundancy to be achieved for each rack, meaning there is one extra PSU per rack of servers (i.e. 5+1) rather than one extra for every server (1+1) which is typical for conventional servers. This reduction in power converting components increases the overall efficiency of the system.
- Oversizing – Conventional PSUs tend to have an oversizing factor between 2x and 3x, whereas a consolidated rack-level PSU will have less oversizing (we assume 1.2).
- More efficient components – The most common grade of PSU used in data centers is 80 Plus Silver standard, which has a max efficiency of 88%. In comparison, data shows 94% and 95% max efficiencies for 12V and 48V rack-level PSUs, respectively (based on data from General Electric). There is also likely to be an efficiency improvement in the VRM when the conversion occurs at 48V instead of 12V.
Findings from the Expert
Here are what I believe the key takeaways are from this analysis:
- Rack-level architectures do show significant improvement in efficiency over conventional servers, primarily due to the reduction in oversizing of the PSUs (both from less redundancy “overhead” and consolidation of server PSUs).
- 48V direct current is a likely bet for the future, not for the incremental efficiency improvement over 12V direct current, but because of the need for increased compute and density.
What do you think? Do you agree with our findings? Will 48V become the future voltage for consolidated power supplies in these OCP-style architectures?
Try Out the Calculator
I encourage you to try the power efficiency tool for yourself. See how sensitive each parameter is to the overall result (move the load levels around, adjust the capacity) and understand just where the efficiency curves fall.