However, whilst the same tolerances of power and cooling are enforced today, the ever increasing load compaction has led floor power densities to rise from 300W/m2 to well over 10,000W/m2. Indeed, individual 600mm wide by 900mm deep IT cabinets can achieve 30kW if fully equipped with the latest blade server technology. This power density has resulted in the need to provide the cooling system with the same amount of reliability, redundancy and resilience as the power system has traditionally been designed to deliver.

The data centre community has its own dedicated trade press and also features in building services and facilities management publications which together provide an informative forum covering technological and management processes, from design & construction through to operation. With so many parties involved it is understandable that you can read about the latest developments in hardware, software, M&E engineering, IT and Corporate Management in diverse trade publications, but in the summer of 2007 the hot topics for data centres were:

• Lower power microprocessors
• Server virtualisation
• Cooling blade servers
• Leading power factor loads
• High availability Tier IV facilities
• Green data centres

Many of these impact the critical load in characteristic, compaction and raised floor power density, but to what combined result?

Microprocessors & servers

From the first microprocessor introduction in 1971 to the introduction of the Pentium IV in 2001 the heat output rose from 2W to 70W. The drivers for the losses were all related to the amount of work done by the chip itself in terms of how much current was switched and how often:

• The number of transistors increased x18,000 to 42m
• The switching speed increased x13,000 to 1.4GHz
• The track spacing decreased by x83 to 0.012 Microns

It is interesting to note that the track spacing is down to such a small dimension that the DC voltage used for the chip has to be <0.5VDC – so that in today’s 140W processor a remarkable 280A flows through the tiny silicon chip at the heart of the processor mechanical package.

Since 2001 the trend has continued and the typical power consumed has risen to 140W. At the same time co-processors and piggyback topologies have been introduced to increase the computing MIPS (millions of instructions per second) capacity in a given volume (MIPS/cm3). The result has been the almost exponential growth in data centre power density (W/m2) and the advent of specialised cooling methods and systems.

The rise of the specialist cooled cabinets and the resultant need for UPS for their mechanical heat extraction systems are current issues for data centre designers.

Against this background some remarkable news has come from the microprocessor OEM’s: The Watts/microprocessor is about to fall from the recent industry norm of 130-140W to 60W or even less over the next 1-2 years. This will, of course, overlay the installed base of servers (with ~3 years hardware refresh time) and, therefore, probably work though to 90% of the population only over the next 4-5 years. It is clear from a review of the technological trends that co-processors (duo and quad etc) and piggyback topology were already accelerating as the power ceiling per chip had been reached at around 140W.

After a little historical reading it was possible to plot and extrapolate the Watts/Chip trend from 1996-2011. Looking back into the historical dimensions (height in U) of servers compared with the numbers of chips installed, it can be seen that the heat rejection and mechanical packages have shaped watts/chip development and that the density rose from around 0.15chips/U in 1996 and can be expected to reach 5chips/U by 2011.

The second factor that has affected the historical power consumption is that of the partial load efficiency of the microprocessors themselves. They have traditionally been very poor in that when processing very little information compared to their capacity the power consumed has been above 80%. This has been improving and it is expected to reach 50% by 2011. To counter this trend the software technique of virtualisation (where processor loads are maximised by the supervisory software) is being adopted very quickly. This is predicted to be fully adopted, with processors running at full load, by 2011.

Taking the exploration of trends to the demand side it would appear from the historical data that the demand for computing capacity per ‘box’ (MIPS) has actually outstripped the MIPS/Chip curve. This is possibly a reaction to the marketing of enterprise software but could also be a negative result of poor, processor hungry software compilation. Taking the demand for MIPS at 50% growth per year but the available capacity per chip at only 30% growth per year produces the trend lines for MIPS (relative to 1996 being 100, over x40,000 by 2011) and the trend for the number of 42U high IT cabinets to house the servers/microprocessors.

OEM’s in the ‘cabinet’ business will adapt to specialised cooling solutions because clearly the number of cabinets is going to reduce even with a huge growth in IT capacity but three interdependent trend results are now sought from the model.

• The total UPS power
• The data centre Watts/m2
• The raised floor area, m2

This model starts in 1996 with a notional 100 cabinet data centre. It can clearly be seen that the W/m2 will continue to rise and, therefore, drive the demand for cooling plant UPS applications.

For the power demand the model produces a very interesting result. It predicts that the IT load will level out (per installation but the number of installations will increase) but that an increasing number of installations will apply UPS to the mechanical load and so drive the power demand upwards.

Printer friendly version

Email this article to a friend