Kvetching about Bitcoin – a quantitative approach
There had been no small amount of consternation about how cryptocurrency and its current uses eats up a lot of power for relatively little gain. However, a Twitter post has made me realise just how little of a number sense I have on this particular subject:
"so mad about the energy consumption of bitcoin" yeah ok how many Chrome tabs do you have open right now
Before we continue further it is critical to note the purpose of this article: it is not to apportion blame as to the amount of waste generated, and it is not to propose any particular solution to that waste, and certainly not to distract from any single cause. The purpose is to, as the subtitle indicates, advocating the practice to assign concrete numbers to problems beyond a human scale, and to gain a sense of number to the problems involved. In doing so it is hoped that one gains an appreciation to how energy expenditure is everywhere and perhaps also to identify other places where such an approach is productive.
Spectacular failures
A few years ago I had the idea of identifying some "spectacular" quality of certain types of waste compared to others. The idea that some forms of negative effects are more spectacular than others, and those that are spectacular tend to have an outsized effect on human decisions.
The classical example is nuclear power deaths: on a joule-by-joule basis, nuclear power kills less than a coal plant. But this is undercut by the fact that when a nuclear plant fails it fails very spectacularly: big doses of radiation, round the clock news coverage. On the other hand, the fact that the primary cause of death with coal power plants is breathing in coal particles and then ruining your lungs is practically invisible and so gets neglected as either bad luck or part of ordinary life.
Another example is plane crashes, which follows much the same pattern: if you had the misfortune to get into a plane accident, it's unlikely to be a pleasant experience. While car crashes kill more, it is possible to have a car crash that amounts to saying "oops" and going to the mechanic's, and this is the experience that most drivers get. So the higher risk of driving is rolled into the risk of daily life and the fear is dismissed.
The central claim for the tweet then is that the spectacular usage of energy in bitcoin may possibly be eclipsed by the much more mundane but still frightening amount used by every Chrome tab open in the world, which by a rough order of magnitude estimation number some 1 × 10↑10 though any particular day and are also known to consume a lot of resources.
(For that estimate, take the number of Chrome users in the world, then multiply by the number of tabs the mean user opens. For the latter, presume that most users open between 1 and 10 tabs, and then take the logarithmic midpoint of that, i.e. 3. Then take the internet penetration rate of the world – 0.57 – and the usage share of Chrome – 0.65 – along with the total population of the Earth, multiply them all together, and round to the nearest half order of magnitude.)
Introducing the competition
Of course, this is all idle speculation unless we can do what was set out to do and put some numbers on some things, and also have some other examples of wastage to compare to. To that end let's introduce some possible candidates.
Web browser engines and JavaScript
It is no secret that web browsers have to do a lot, and if it is so that the greatest use of CPU cycles in a typical human-operated computer is to render web pages and execute JavaScript I would not be surprised. With so many web browsers around the world, it's reasonable to presume that they represent a significant chunk of global computer energy usage.
To properly quantify this usage, we need to consider the amount of browser instances and JavaScript interpreters open in the world, then estimate their power draw (in watts). Then a simple multiplication of the mean lifetime of the instances and interpreters will give us the total energy draw.
One can estimate power draw fairly easily by monitoring a typical Facebook session and counting CPU cycles and therefore also FLOPS/W (which reduces to FLOPS/s because the seconds cancel out – the abbreviation invisibly changes from "floating point operations per second" to just "floating point operations"). Facebook is chosen because it is most representative of the mean usage pattern of the typical Internet user. As for the mean lifetime, we mainly presume that most tabs live essentially forever, so we choose 1 day as the lifetime so as to match up with the assumption above about the average tab count in the world every day.
Advertising technology
Later in the replies to the tweet above, a new challenger approaches:
the ad tech cat-and-mousegame bullshit on media sites probably eat bitcoin for breakfast
This partially overlaps with JavaScript usage and the remainder is mostly typical tech company power draw, but consider that advertising on the internet in particular is one of the most wasteful ways that one can spend resources of any kind prior to the proliferation of NFTs.
One particular image that has been impressed on me, of which I have misplaced the source on, is that most of the advertising ecosystem is extremely lossy due to the large amount of fakes involved in every level: starting from scam vendors, to fake agencies, to opaque auctioning procedures, to bot audiences causing any ads that do make it to land on nowt but a smartphone farm somewhere in China where they are lined up in display racks with robot arms tapping and clicking away. Humans also have ad block, reducing the efficiency even lower.
It is therefore a reasonable proposal to split out advertising technology into its own category as being particularly wasteful and unwanted, even though it might also be comparable to typical computer usage of most technology, research and finance companies, because they are all roughly stereotypical office enterprises with a data centre and a focus on computation.
Quantifying this is a bit more complicated but doable with a little bit of extra knowledge that should be accessible. The key parts are:
First, determine what every adtech company is. There are a number of photos floating around that show the dizzying array of unfamiliar names and mysterious data flows in this sector so this quantity is within reach.
Second, determine the average tech and energy usage of these companies. For this we stereotype them as above and find the rough electric bill that a representative company has, possibly from its tax return along with the tax return of whatever office it is renting. Then we can subtract away lighting, cooling and heating from physical arguments such as floorspace and office hours. Finally, for the more forward thinking companies that have many employees working from home, we can add on the domestic electric bill of a typical household, per capita, multiplied by the number of employees with a fudge factor as needed.
One can also additionally estimate the end-user impact of adtech by using the same experimental setup as above, but this time analysing the JavaScript usage by origin domain and identifying those that are related to advertising. This allows us to separate the advertising usage within JavaScript to everything else and we can then apportion this amount to adtech, JavaScript or both as desired.
Other IT office work
Since we mentioned technology, research and finance, we will have to include this as a particular candidate for wasteful energy usage. This is contentious depending on the particular audience, but some would consider some of these fields to have no value to them or society at large and therefore be wasteful. Regardless of any particular stance, it is helpful to calculate the energy consumed in this case anyway so as to get more perspective.
The approach for these is essentially identical to the adtech example given above, as we have stereotyped their activities to be also the same. The only difference is that we need to have a list of firms and a classification of whether they belong to any of technology, research, finance, or some combination thereof.
Old hardware
In contrast to the above categories, which are mostly about the quantity of computation, this is more about the amount of quality that the computation rests upon.
While every year the amount of energy used per CPU and GPU cycle continues to decline, there are still many places that use old hardware, either out of necessity or simply because it is similarly wasteful to upgrade. Regardless, not using more efficient hardware still carries with it an energy cost that is continuous and constant.
This is the trickiest quantity to estimate of them all, partially because it is qualitatively different and there are some difficult mitigating measures that stretch beyond mere computation. I propose the following approach:
- Enumerate all processors.
- Identify which equivalent is the most energy efficient.
- Calculate the power draw difference between them under some representative load.
- Subtract away the cost of upgrading to that most energy efficient processor, including logistics and assembly.
The last point is what makes this computation particularly hard. I do not presume that this is a significant competitor, but it is sufficiently widespread that I must mention it. A full work would probably omit this partially on feasibility grounds and partially because of the amount of work required against what would result from it, but including this will help us gain a sense of perspective of how much waste we should actually expect, given that old hardware is a fact of life that will happen no matter how many revolutions you throw at it.
So are you going to do the maths?
No.
As outlined here this is already a significant research project that would take a number of months, and I banged this out during a lazy Saturday afternoon while thinking about energy usage over lunch. The amount of work required to actually compute these quantities is far too large for your typical Breakfast Experiment capacity.
Second, astute readers might notice that this is essentially just an outline of a Fermi estimation of energy use. Fermi estimations are powerful, but they are even more powerful if we can consider multiple ways of reaching the same quantity. I've only outlined one here, so this is not the optimal way of finding the required numbers. Fermi estimations also have the flaw of missing out on certain factors that may not show up during dimensional analysis, and this problem in particular has a lot of dimensionless numbers that are invisible to such analysis. A serious computation would need to ensure that the methodology is mostly correct, and that is beyond the scope of this article.
So what is the point here? It is to gain a sense of perspective, and in this case the best sense of perspective is to give an order-of-magnitude estimate. In this case what we want to know is whether or not bitcoin energy usage is unusually large compared to "ordinary" energy usage with computers, so if it turns out that mining cryptocurrencies are 10× or 30× or even 100× larger than other computations this analysis will show it. On the other hand, if it's roughly comparable in order of magnitude, like 0.3× or 1× or 3×, then maybe we need to reconsider the fervour against mining as even a unique devil in wastage.
Doing the calculation as outlined here will give you that perspective. The vagaries involved would mean that ultimately the value is only usable to the nearest half-order of magnitude, that is, a number that starts with either 1 or 3, and the rest of the digits are 0s. Order-of-magnitude gives you a good sense of basically anything in this universe, as it allows us to have a reasonable ruler that stretches from minute to enormous, and having this is the key to understanding the relative size of everything. If for whatever reason this computation is to be carried out, I believe that one would gain a similar understanding to energy use, which thus far has been discussed mostly in isolation or in pairs, without knowing the entire scale of consumption.