Predictions and Forecasts?
Most of us like a bit of speculation and making lists so this month, when asked for ‘predictions’, I have decided to make a list of my top three.
The first in my list, ‘in no particular order’, concerns ASHRAE and their Thermal Guidelines for microprocessor based hardware. We, the data centre industry in Europe, are very lucky to have ASHRAE. OK, they are a purely North American based trade association that serves its members but they are the sole global source for the limits of temperature, humidity and air-quality for ICT devices and have proven themselves to be far more progressive for the environmental good than anyone could have expected. If you follow their guidelines from the first to the latest you could be saving more than 50% of your data centre power consumption since the members of TC9.9, who include all the ICT hardware OEMs, have consistently and regularly updated their Thermal Guidelines, widening the temperature/humidity window to enable drastic improvements in cooling energy.
So, what is my prediction? Well it certainly is not that they will make the same improvements in the future that they have made in the past, since the latest iteration leaves server inlet temperature warmer than most ambient climates where people want to build data centres and requiring almost zero humidity control. My prediction is, in fact, that the conservative nature of our data centre users will keep the constant lag in ASHRAE adoption at a lackadaisical and slightly unhealthy five years. What I mean by that is simple – the 2011 ‘Recommended’ Guidelines are, in 2017, just about accepted by the mainstream users as ‘risk free’ whilst many users still regard the 2011 ‘Allowable’ limits as avant-garde. So, I predict that ‘no humidity control’ and inlet temperatures of 28-30°C will be mainstream by 2022…
The second prediction in my trio concerns the long-forecasted, but now clearly closer, demise of Moore’s Law. When Gordon Moore, chemical engineer and founder of Intel, wrote his Law it was clear to him that the photo-etching of transistors and circuitry into silicon wafer strata doubled in density every two years. That was soon revised by his own company to a doubling of capacity every 18 months to consider the increasing clock-speed and, more recently by Raymond Kurzweil (sometime nominated as the successor to Edison), to 15 months when considering software improvements. It lost its simple ‘transistor count per square mm’ basis long ago but Koomey’s Law took up the baton and converted the 18 months’ capacity doubling to computations per Watt. Effectively that explains why it is so beneficial to refresh your ICT hardware every 30 months (or less) and more than halve your power consumption for the same ICT load. To make a little visualisation experiment in ‘halving’ take a piece of paper of any size and fold it in half, and again, and again... You will not get to seven folds since you will have reached the physical limit.
So why have data centres been growing if Moore’s Law and its derivatives have been providing a >40% capacity compound annual growth rate (CAGR)? The explanation is simple – our insatiable hunger for IT data services (notably including social networking, search, gaming, gambling, dating and any entertainment based on HD video such as YouTube and Netflix et al) has been growing at 4% per month CAGR (near to 60% per year) for the past 15 years. The delta between Moore’s Laws’ 40-45% and the data traffic rise at 60% gives us the 15-20% growth rate in data centre power. The problem comes when Moore’s Law runs out, which it surely will with a silicon base material, as then we will have to manage the 60% traffic per year without any assistance from the technology curve.
Moore’s Law probably has 5 years left without a paradigm shift away from silicon (to something like graphene) but that is unlikely to happen ‘in bulk’ within the 5-year time frame. Looking at one of the major internet exchanges in Europe shows that peak traffic is running at 5.5TB/s with reported capacity at 12TB/s – but if we consider even a slight slowing down of the annual growth rate to 50% then it will be less than 2 years before the peak traffic will be pushing the present capacity limits. I predict a couple of years of problems during the dual-event of a paradigm shift away from silicon and a sea-change in network photonics capacity.
The last of my trio of predictions concerns the reuse of waste heat from data centres and is simply stated as: By 2027 waste heat will not be ‘wasted’ from a huge array of ‘edge’ facilities and they will become close to net-zero energy consumers. From my perspective, there is a gathering ‘perfect storm’ of drivers that will converge to drive infrastructure designers to liquid based cooling:
Data growth will continue to exponentially grow driven by UHD-video (4K followed by 8K) and The Internet of Things
Data centres will be at the heart of the ‘smart’ everything; cities, grids, buildings and transport
Huge ICT/kW, despite a bump-in-the-road for Moore’s Law, per Watt will continue to grow at 4% compound per month and so will reach 80x today’s capacity/Watt by 2027
ICT load will (with market saturation and slower population growth) tail-off to 3% compound per month but still with a gap between generation and capacity of 1% per month. That represents a data centre power growth from 3% of the utility today to 20% by 2027 – with ICT in total reaching 60% of the utility, a totally unsustainable proportion in a period requiring drastic power saving and a change to renewables
Air cooled hardware at 27°C inlet temperature struggles to reach 35°C exhaust which is very low-grade heat and can hardly be used economically without complicated and only partially efficient heat pumps
The solution is simple and within our grasp today – liquid cooling of the heat generating components, particularly the microprocessors. With liquid immersed or encapsulated hardware and heat-exchangers pushing out 75-80°C into a local hot-water circuit with 94% efficiency the data centre will have a net power draw of just 6%. Just five cabinets (a micro-data centre by todays definition), equivalent to 80x todays ICT capacity, will be able to offer to the building 100kW of continuous hot-water. Consider embedded 100kW micro-facilities in offices, hotels, sports centres, hospitals and apartment buildings. Indeed, could this be the ‘major’ future? Could giant, remote, air-cooled facilities become obsolete? Probably not for twenty years, but then…