Powering the infrastructure of the cloud

Interviews |
By Nick Flaherty

Cloud computing has become a signficant driving force for power management technology. The latest generation of high performance processors, FPGA and AI acceleration engines is requiring many more power rails that need both tight control and synchronisation. As data centres upgrade their compute engines the power requirements are becoming increasingly challenging.

This has been a key area for Analog Devices (ADI) following the acquisition of power specialist Linear Technology in July 2016. One of the technologies that came with the deal was a family of micromodules to provide DC-DC conversion at the point of load (PoL) which is becoming essential for the transition to power system management (PSM).

“We are right now on the front edge of realising our fourth generation technology,” said Chris Mann, vice president of the power products group (above left) . “For example our 12V point of load step down converter delivers 100A at 1V output. This is a power density 2.5 times our original parts.”

“Higher current is clearly a target for us as customers are asking for hundreds of amps and at the moment that’s multiple modules – 100A is not the limit,” he said. But it’s not just about the overall power levels that are important but how that can be delivered on the board. 

“Increasing power density is important but where that power density will reside on the board is important,” added Rob Chiacchia, general manager of Power Systems in the Power Products Group and responsible for the micromodule activity (above right).

“Although we are putting components on top to increase the height and reduce the footprint, we are making components thinner so that they fit on the back side of the board to alleviate congestion on the front side that our customers are running into – if customers can put component power on the back side of the board their layout is easier,” he said.

“Packaging is key – we are running out of space on the board. The 4644 quad output is in an overmoulded package but the components are all on the top layer of the laminate – we’ve been asked to increase the power of the outputs and instead of increasing the footprint we are using the component-on-package technology to take the inductors off the laminate and that gives us more space on the laminate for switch and boosts the airflow with better thermal capability.

This takes significant optimisation he says. “We are working on four different designs, and each time we do that requires another inductor design. For example the latest generation 4700 is a different inductor design designed with component-on-package in mind.”

“Another direction is to go from 12V to 48V and when you do that you reduce the input losses by a factor of 16 so high power systems that need higher efficiencies are going to higher bus voltages. Yet another direction is 48V to core topologies that allow 48V to core voltages at lots of current.”

ADI has also been looking at different conversion topologies such as switched capacitors.

“One micromodule is based on a switched capacitor front end that very efficiently converts 48V to 12V and then to the core voltage very efficiently – that’s a controller product line developed at ADI,” said Chiacchia. “What’s different is the switch cap front end is 98% efficient at high power levels. We have another topology based on a full bridge primary transformer that is very efficient in delivering large currents to these core voltages. There’s a lot of activity in the market in this area.”

The evaluation board for the LTM4700 micromodule 

Wide bandgap devices

While there is a lot of focus on gallium nitride (GaN) and Silicon Carbide (SiC), the power division is still focussed on mainstream silicon based around its high speed SilentSwitcher architecture says Mann.

“We are certainly interested in GaN for higher efficiencies,” he said. “In the 12V space we think that silicon still has major advantages over GaN and has room to run but we do see SiC and GaN being great for higher voltage conversions above 60V. I don’t think we are focussed above the 48V or 54V rail.”

Other parts of ADI such as automotive are looking at 800V to 5V topologies though to get the higher switching speeds.

“We are switching faster every year and that allows us to reduce the size of the magnetics and puts different requirements on the power stages,” said Chiacchia.”We find we have a huge advantage with SilentSwitcher with very fast edge rates to switch faster without the inherent switching losses so going forward we expect to use this in our micromodules to use silicon FETS for higher frequencies –

“It has a lot to do with the current levels,” said Mann. “5MHz at 5A is possible today but moving an order of magnitude to 50A is achievable.”

“We do have a new product line coming out that are smaller than the original modules at 6.25mm on a side with 20A so there’s a market for that general purpose micromodule power that you can sprinkle around the board.”


Another key area is communication between modules. With distributed power, balancing the requirements is essential. PMBus is seen as a high end feature on power supplies, but there is a move to use it more in individual components to enable lower cost telemetry, margining, supervision, sequencing and fault logging in large power systems.

“We do hear that customers want lower cost implementation of PMbus,” said Chiacchia, who was one of the original directors of the group that put the specification together.

 “When PMBus came together, our goal was to try to make firmware reusable in complex power systems that needed complex sequencing and complex margining testing and provide programmable monitoring to detect faults but you have to manage those faults,” he said. “Our devices include a non-volatile fault log that records the faults without the system having to go down.”  

The modules use a proprietary one wire interface as device synchronisation and chip communication was intentionally left out of the PMbus spec, which is currently at version 1.3. This was to encourage technology development among the suppliers, he says.

“You will see solutions that go both ways. We introduced a low cost simple PMbus 1.3 compliant DAC – it’s just a D to A converter that allows you to sum a current into a regulator but it complies with the high speed bus – that’s a very simple device, it doesn’t need to be configured and it’s low cost. You will see lower feature count, lower cost telemetry where instead of monitoring 20 parameters we monitor 4 – input current and voltage, and output current and voltage,” he said,

But there are also demands to support high performance multicore processors and FPGAs in datacentre infrastructure. This is leading to a new version of the specification.

“On the high end we see customers that need to power very complex, power hungry FPGAs or cores with precise control of the DC transient specifications and in addition they would like the system to be very adaptive with an almost infinite number of power states and for that they will need. They will need 1.3 with the high speed 50MHz SPI  bus – this represents the other end of the spectrum that needs to communicate lots of information to the power system,” he said.

“So 1.3 is very much like the VR13 and VR14 specifications from Intel that does essentially the same thing but PMbus is more about system level issues than just the core voltage,” he said.

“An update to version 1.3 is in the works to accommodate multiple slaves on the bus – both Intel and PMbus will be changing the specification to provide multiple rails and not just the core power supply. That wasn’t envisioned in the original spec so a protocol change is being developed right now,” he added.  

Power over Ethernet

Another area that is of growing interest is Power over Ethernet over Category 5 cabling to support the roll out of smaller basestations for 5G cellular networks.

“We are the first to market with a 802.3bt 71W standard based on the PoE+,” said Mann. “Linear Technology was always more forward thinking and always coming up with formats for higher and higher power.”

“The power is really limited by the resistance of the cable so we are working with customers on delivering 150W to the load through PoE++. This is a trend that we are seeing in small cell applications that are powered directly by the Cat 5 cable. A lot of these installations have fixed cat5 or cat6 cables with lower resistance that allows more power to the load. What we see is that small cells will continue to grow,” he said.

“We are seeing a rapid adoption of the ideal diode bridge controllers that reduce the power loss in the PoE bridge at the end of the cable prior to delivering the power so when you reduce the loss of the diode bridge using a MOSFET and controller you can save thermals and reduce heatsinking. But for that you need PoE++.”

But customers also want the cells to be backwards compatible with all the specifications to power devices using 13W, 25W, 71W or 150W for the different PoE specifications and this is driving the adoption of the latest PoE++ specification to ensure that all kinds of cells can be powered by the infrastructure.

This is not just driven by the move to 5G, but an increasing interest in satellite comms as well. More small nanosats are being placed in orbit to provide communications links.

“The 5G small cell activity is mainly Asia but we are also seeing other opportunities in satellite comms where a terrestrial basestation in a home would have equipment that needs to power an antenna on the roof and the connection is through ethernet to carry both data and power. The power requirement is in the 100 to 150W range so we are also very busy in satcomms and that’s in the US,” said Chiacchia. “We think that’s going to be a high volume application with Internet from space.”


Linked Articles
eeNews Power