PROBLEM /CHALLENGEBlade server computing systems must be highly configurable to meet a variety of application needs. As a result, blade servers are bought and sold primarily based upon a
desired amount of processing power required to meet the needs of a specific task. Ironically, one of the least considered, and yet most important, components of a blade
server is the power supply. After all, it can be easily argued that a system¡¯s reliability is only as good as the power supplied to it.
Texas Instruments recently worked with a customer wanting to design a 1.5kW, highly-efficient, isolated power converter module for a blade server application. This customer
relies heavily upon a wellknown power module supplier for meeting his power needs. Power converter modules offer many advantages, but the foremost must be ease of
implementation.
However, there are several disadvantages with this approach, particularly in blade server applications. Server systems are designed to be highly-configurable in terms of
available processing power.
To handle a variety of scalable power requirements, a configurable server needs to be optimally matched to a compatible power system. As a primary obligation, the customer
desired more flexibility by closely matching their power needs and processor requirements.
Secondly, because of specific supervisory and power management requirements, the power modules themselves do not make up a complete ¡°power system¡±. And finally, the
customer emphasized the need to reduce overall power system costs.
Faced with a certain cost increase associated with outsourcing a custom power converter design for their blade server applications, this customer boldly decided to do the
unthinkable—design his own custom power supply.
SOLUTION:Traditionally configurable servers have been built based upon an industry standard rack capable of holding up to 42, 1U (1.75 inches) servers. Since AC power is bussed
through the server rack, eachhorizontally-mounted 1U server module requires one isolated AC to DC power supply, or two if redundancy or load sharing is required, and its own
cooling source.
Consequently a fully configured 42, 1U rackmount server can house as many as 84 individual power supplies.
Blade servers are mounted vertically into a 3U rack unit. Each blade is really nothing more than a single board computer, complete with memory, processor and network
connections. By sliding the blades into a common vertical enclosure, a larger number of servers can fit into a smaller allowable space resulting in higher computing power
density compared to horizontal rackmount servers. As many as 250 blades can be installed into a single PCI based industry standard rack.
Unlike the 1U horizontal rackmount servers, DC power in ablade server is distributed from one or more power supply modules, through the PCI cabinet and to each of the
individual blades. Distributing DC power means that the AC to DC isolation only has to occur once through an intermediate buss converter. Since the blades themselves do not
have an on-board isolated power supply or fan, buss power and cooling distribution can be thought of as shared system resources. Blade server power supply modules can now be
paralleled and hot-swapped for greater system flexibility and scalability. In their earliest development, blades contained only a single processor so power dissipation was
not a serious issue.
However, the current trend is to pack as many as four power-hungry processors onto a single blade which greatly increases cooling demands. And because the blades themselves
are now dissipating so much additional power, there is an everincreasing demand to push for higher power supply efficiency.
Because the customer specializes in high-performance computing systems with little to no previous power supply design
experience, designing an application specific, custom power supply was a significant challenge. TI worked closely with the customer to first understand their immediate power
requirements. The 1.5kW converter needed to operate from an off-line, power factor correction (PFC) regulated boost converter. Initially, the UCC38500 combination PFC/PWM
controller was considered, however this approach was eliminated because it was limited to several hundred watts of power. A version of the full-bridge power topology was
really the only practical choice and for maintaining absolute highest efficiency, the UCC3895 phase shifted full bridge (PSFB) PWM controller was chosen along with a UCC3817
PFC boost preregulator.
The PSFB control technique differs from the traditional PWM full bridge in two ways. First, an intentional delay time is introduced where, under the right operating
conditions, zero voltage switching (ZVS) occurs, which would be essential for meeting the customers need for high efficiency. Secondly, the duty cycle of each primary bridge
MOSFET is fixed at 50 percent so instead of the pulse width being control modulated, it is the phase shift (or phase overlap) that is controlled.
For a comparable component count, the PSFB offers ZVS, EMI and performance benefits unmatched against the traditional hard-switched full bridge technique. Optimizing the
converter for high efficiency depends upon a thorough understanding of parasitic circuit elements such as MOSFET junction capacitance and transformer leakage inductance.
Even once these non linear elements are defined, the ZVS critical timing must be precisely adjusted to fully realize the converters¡¯ efficiency benefits. Leveraging TI¡¯s
expertise in ZVS power topologies, the burden of climbing an otherwise steep learning curve was avoided. Throughout the entire development cycle, TI provided extensive design
support with specification and design reviews, component recommendations and support of initial hardware debugging. Within nine months, the customer had its first ¡°in-
house¡± custom designed power system ramping into production.
A short time later, this customer again consulted TI to discuss their specifications for a 200W DC to DC converter
design. Since favorable results with the 1.5kW PSFB design had been achieved, the customer naturally considered using a scaleddown 200W version. The PSFB would undoubtedly
meet their requirements of high efficiency and reduced cost, but was it the best choice? The singleended forward converter with active clamp transformer reset offers many of
the same advantages found in the PSFB topology, but at reduced power levels. Where the PSFB can be designed to operate at tens or even hundreds of kilo watts, the active
clamp forward is limited, in a practical sense, to about 400W. Once the performance and cost trade-offs were presented, the customer agreed to proceed with an active clamp
forward design based upon the UCC2894 Active Clamp Current Mode PWM controller.
To help quickly bring the customer up to speed, TI provided a reference design and a working converter used as a learning aid for gaining a thorough understanding of active
clamp forward converter principles. TI also delivered a customdesigned mathematical model to predict the power converter module efficiency. Efficiency modeling is useful
because it speeds up the design verification process by allowing exploration of the effects on efficiency while varying design components or operating parameters. With TI¡¯s
design support, this customer was able to take their first active clamp forward converter from concept to preproduction in less than six months.
The customer now has experience with two of the most efficient power topologies available and has scalable power solutions from less than 100W up to many kilo watts. In
addition, they have acquired a real appreciation for all that is involved in custom designing efficient power conversion products for their blade server systems. The customer
has even been discussing the possibility of allocating in-house design resources uniquely dedicated to developing custom power systems for blade server applications.
