The engineers at Microsoft Corp.'s Global Foundation Services unit (GFS) are among the professionals looking to tame the data center's runaway power requirements into submission. GFS distinguished engineer Dileep Bhandarkar talked about the problem and potential solutions with EE Times semiconductor editor Mark LaPedus.

EE Times: What is the overriding goal within a data center?

Dileep Bhandarkar: The overall metric is the total power consumed per unit of work. This includes efficient cooling and power distribution, judicious allocation of power, increasing the utilization of the server hardware, and better power management. Packing more useful work within the total power budget of a data center is our goal.

EE Times: In terms of chip design for servers, what are some of the problems?

Bhandarkar: The main challenge is power. At Microsoft, we work with our suppliers to remove unnecessary components and use high-efficiency VRMs [voltage regulator modules] and power supplies. We recently decided to use low-power processors in our standard servers, given that you can save 20 watts per processor if you are willing to sacrifice 5 percent in frequency at the same price point. If our servers are not heavily utilized, the resultant 2 to 3 percent decrease in performance is acceptable. As multicore provides higher performance, we also see an opportunity to replace some of our four-socket platforms with two-socket platforms, and use single-socket in place of two-socket. Balancing and right- sizing the configuration are the answer.

EE Times: For DRAMs, is it a simple process of lowering the voltages from 1.5 volts to 1.35 V?

Bhandarkar: It is more than just lower voltage. DRAM vendors also need to work with microprocessor vendors to make sure that their devices allow the memory controller to implement power-saving techniques.

EE Times: Do you see the data center world going to more-traditional DDR SDRAMs or perhaps to a nonvolatile approach, such as EcoRAM, phase change, MRAM, etc.?

Bhandarkar: This is an issue of technical feasibility and economics. The evolution of existing technology often has more improvement, and revolutionary approaches take longer to become economically viable. We will adopt the best available commodity technology.

EE Times: Will solid-state storage cut the mustard for storage, or will disk drives prevail?

Bhandarkar: SSDs will find use in places where its high bandwidth is needed while low capacity is sufficient. We do not see SSDs displacing hard drives across the board anytime soon. A combination of the two where the SSD is used as a cache for disk drives is a promising idea. This is similar to how SRAMs and DRAMs have coexisted for main memory.

See also: Green-memory movement takes root