Is data centre power going to fall? The answer to this question, is: ‘probably not just for the moment’ but it certainly looks like the exponential growth in data traffic that we are used to reading about has started to slow down. It has not actually gone down yet, but the growth rate in markets with high mobile phone penetration, or businesses not based on HD video and images, has shown a noticeable hiccup for first time in nearly two decades.
That is the problem with using hyperbole like ‘exponential growth’.
Compare the results after 10 years of some example compound annual growth rates (CAGRs). A healthy business might achieve 10% CAGR in profits, which would produce 2.6x after 10 years, while Moore’s Law (to which we shall return) at 40% CAGR would reach 29x and historic data growth at 60% CAGR will grow by 110x – all ‘exponential’ but certainly not equal in excitement.
Now data generation, transmission, storage and dissemination take energy, so the growth in data traffic can serve as a proxy for growth in power consumption.
On the other hand, we know that Moore’s Law has been the main driver behind the exponential compute capacity of the ICT hardware: firstly by Gordon Moore (more than 35 years ago) predicting a doubling of the number of transistors on a single chip every two years; then revised by his company, Intel, to doubling every 18 months to allow for the increased clock-rate; and, finally, revised by Raymond Kurzweil, who combined the physical attributes with better software and came up with a doubling in capacity every 14 months.
More recently, Koomey’s Law has elegantly re-expressed Moore’s Law in terms of ‘computations per kWh’ and showed the trend back to the earliest powered computational machines, rather than the 2,600 transistors per chip in 1971 that Gordon Moore started with. The result can be best expressed by considering a facility built in the year 2000 with 2,000 cabinets rated for 1kW each. It can now be out-computed by a single 5kW cabinet – a staggering technical achievement by any measurement or benchmark. It has certainly made any improvements in PUE paltry to the point of almost inconsequential.
So, given the huge capacity curve, why do we have so many data centres and why have they grown in power rating? We can simply compare the two growth rates, data versus capacity. For the past 17 years (since June 2001), the monthly data traffic statistics of one of Europe’s largest internet exchanges, Amsix, have been published (see Figure 1).
The growth until the beginning of 2017 followed a fascinating yearly cycle, with most growth in the first six months of each year followed by a relatively stable platform – but always clinging on to an exponential curve of +4% compound monthly growth-rate. That is +4% a month, every month.
Starting in June 2001, the input traffic was 690TB/month but by December 2017 it had risen to just over 1.2 million TB/month (with a record peak of 5.6TB/s) – a 1,700 times increase and equivalent to a 60% CAGR. The ‘staircase’ that you can see in the growth curve is fascinating, with ever higher ‘rise’ and steeper ‘going’ – until January 2017. The dotted line is +4% per month.
Over the same time, the ICT hardware, from Moore’s Law et al, has followed at least a 40% CAGR (the red line on the graph), so the delta between demand and capacity, 20% CAGR, can be used to explain data centre growth.
It is also not wholly unreasonable to attribute most growth to the individual national penetration of mobile phones/PDAs, high-speed networks and high disposable income. Faster broadband drives energy demand. Hence markets with a low birth-rate, high disposable income, ever-fast access enabling lots of social networking, HD and UHD/4K video entertainment and >100% active SIM cards (compared to population, such as the UK with 20 million more SIM cards than people) will be experiencing a slowing of the data growth rate compared with, for example, an emerging economy with increasing proportion of younger and aspirational population, rapidly increasing mobile phone coverage and device ownership which is skipping copper in favour of fibre and going straight to 5G – where the current annual growth rate will be considerably higher than the 20% suggested above.
Certainly, I know of several enterprise facilities, whose business is not predicated on HD images, gaming or gambling, where their ICT service is increasing in volume but their data centre power is declining annually as they refresh hardware.
The faster they virtualise and refresh hardware (maximum three years, better two years) the faster their data centre loads are tailing off. And how many of you have experienced collocation clients whose ‘growth plan’ has been exceeded lately? In my experience, quite the reverse is common.
Because of these two diverging growth curves I have been predicting for the past 10 years that the demise of Moore’s Law (which, if not already dead in its original meaning, will certainly not last much longer with silicon as a substrate) will lead to a capacity crisis as the tsunami of data overwhelms the fibre networks. OK, a change from silicon to graphene may postpone the problem for a decade or so but the idea that the hardware will cease to follow its historical capacity per Watt has been alarming, to the least.
I have been known to often quote Vint Cerf – a personal hero of mine, inventor of the IP address and one of the ‘fathers of the internet’ – who said in 2012: “Internet access will become a privilege not a right.” Of course, there other things we could do to reduce power – utilisation being the low-hanging fruit – but we should do those regardless. So, why do I now think that the future looks less apocalyptic? Please forgive the hyperbole but the first way of looking at the monthly data is to compare the year-on-year growth and we can immediately see a trend of slow deceleration. It may be a premature reaction but the past 12 months have broken the pattern that has been so resilient since the Amsterdam Internet Exchange monthly data log started in 2001. Just compare the monthly input traffic of 2006 with 2017, just a decade apart (see Figure 3).
The monthly traffic has largely been static through 2017 with much less variability per month than ever. Starting the year in January with 1,174,500 TB/month and ending in December with 1.25 million TB/month (only a 6% increase for the year), and not setting a new peak every couple of months (as has been the norm in prior years), is a remarkable change. I’m feeling optimistic, maybe misplaced, but it will mean that data centres will continue to grow in the emerging economies where phone penetration is not yet >100% but elsewhere it may stabilise.
Now, 6% CAGR would be healthy for most industries, but it is far less than the exponential explosion which went every year before. Maybe Moore’s Law and all its iterations will finally start to overhaul the demand? At least in Europe, and I suspect in the US well before the emerging economies.
It could well be that European traffic is slowing down (not, yet, ‘going’ down) as high mobile phone penetration reaches into every corner of the poorer parts of Europe?
It may depend upon the Internet of Things and ‘edge’ computing but if the content is kept local (and maybe if the content isn’t UHD/4K) the growth in ICT hardware and photonics research should stay ahead of the roll-out – which will, in any case, start in the richest parts of the world and make the gap between the rich and the poor even wider.
As an aside, it is rather odd to see the human need for a ‘better’ phone – it’s not unusual to see the latest smart-phones in regions of South America that are still working on 2G.
So, what about January and February 2018? I added January’s and February’s data. The stagnation in monthly traffic continued (and actually reduced) and the gap between the (16y) traditional 4% CMGR and 40% CAGR (Moore’s Law) curves dramatically widened.
The European traffic has fallen but it might be a blip on a bubble and all kick-off again… I’ll get back to you later with an update of war or peace breaking out.