Why Do Computers Use So Much Energy?
blogs.scientificamerican.com
blogs.scientificamerican.com
30% goes on transport.
> the human brain is a computer. This particular computer uses some 10–20 percent of all the calories that a human consumes
So society's computer uses 5 percent and our own, evolved over millions of years, uses up to 20.
I'm not saying computers can't or shouldn't be made more efficient, but relatively speaking, the title of this article is pretty bizarre. Computers don't use very much energy compared to almost everything else in the world.
> But the thermodynamic efficiency of this computation—the amount of energy required by a ribosome per elementary operation—is many orders of magnitude superior to the thermodynamic efficiency of our current artificial computers.
So I think the headline is justifiable in light of this comparison.
Computers are so much more efficient that so many things in our world that we could much more easily improve the efficiency of: it would be great to make artificial computers as efficient as our brains but sinking resources into the effort rather than addressing much more pressing low-hanging fruit isn't really logically sound.
Of course it's not a zero-sum game: we can sink resources into both, but the alarmist headline is still hyperbole for this reason.
Also, let's not forget about the enormous time, space and energy used by evolution to make humans. Nothing we invented can rival that kind of resource usage.
Even in Deep Learning, performant models require more data and compute. You can't obtain the same accuracy from running a computer for a day compared to 500 TPUs for a week. Humans have had lots of time to train and evolve, computers not.
And as long as we can't get our energy without killing other living organisms, the energy brain uses can't be compared to the energy computers use. It's physically the same, but morally not. Of course you could say computers are here because humans have killed living organisms to feed themselves, so computers partake into that, but we'd end in a loop and never reach any conclusion.
It is perfectly possible to fuel a human only from biological sources that have died a natural death. It's exactly the same as running a computer on fossil fuels, just without the millennia of waiting for sufficient accumulation that you need for building a whole industrial base in one short flash in the pan moment (compared to the millennia of accumulation).
Actually, you don't even have to wait for organisms to die, you can even use fuels where the providing organisms live on, having just shed off a part they don't need anymore. No killing involved in eating an apple. And if you plant the seeds in an advantageous environment you are even taking an active role in their procreation, as the seeds that just drop into the shade of the parent tree won't thrive there.
The computer is very efficient, electrical engineers toil and sweat a great deal to make them so. Then we build things with Ruby and Python and make things 2 orders of magnitude less efficient for, I think, no good reason. What makes those languages pleasant to use is not strictly tied to them being interpreted.
If it takes a developer hours to build something that would otherwise take days, that's a very good reason. Processing power and energy are a tiny fraction of the cost of dev time. For things where scale matters, companies switch to faster languages with more optimization. This is irrelevant for what, 95% of businesses? Granted, probably half of devs are working for big tech where it does matter, but it really doesn't for a lot of us.
In any case, Javascript, which most of us are using more and more, is not that much slower than Java or other low level languages and can be used everywhere.
I think that is in a lot of cases not true, just that the developer (and his company) is not paying the bill but the buyer/user. While individually this might not be a lot, therefore the individual will not complain about it either, for applications with lot of users it will add up a lot. Let's say you have a application or a device with 1 Million Concurrent users, and you could reduce the power consumption by just 1 Watt per instance, this amounts to nearly 9000 MWh yearly, which as a total cost for end users in germany would amount to more than 2,5 Million Euro (+ the climate emissions). I am quite sure that in a lot of instances you could reduce the power usage by 1 Watt and that it would cost less than even 10% of the aggregated savings. Even simple things like game menu screens having no frame limiters and putting the GPU under heavy load have been a thing in the past (e.g. Diablo3 at some point). The issue is that the developer has no incentive to improve such things, except in the most extreme cases as he doesn't benefit from it but from a macro-economic and ecological perspective it would be beneficial.
Javascript is used client side, and again, it's quite fast, with performance comparable to Java and C++ in most use cases. And a web app is basically never going to reach the cpu/gpu levels that even the lightest game coded in C++ will reach, with rare exceptions.
It's a trade-off. You can compute everything in C, or some things in python. In the former scenario you're trading off software engineering cost with computational efficiency. Since software engineering is insaaaaaanely expensive (I code stupid stuff all day and am paid $150k/y) and processors are dirt cheap compared to them, it makes a lot of sense to do some computing in python/ruby/javascript.
One of them is over the total energy consumption of the society and the other over the energy needed be keep a human alive.
Computers would not get "magically" more efficient if another industry triples to overall energy consumption of our society, yet that is what your analysis would suggest...
In other words: There are plenty of things in the world that use an awful lot more energy than computers.
That fact doesn't make computers any more efficient.
This comparison doesn't make a lot of sense to me. It's not even comparing apples to oranges, it's comparing apples to the concept of taste.
Furthermore these percentages don't really tell you much regarding efficiency, I'm sure there are less developed societies where computers count for less than 5% of all energy consumption, doesn't mean that their computers are more efficient.
Also, your emissions are ultimately plant based, so the only real net emissions are due to producing fertilizers and machinery for agriculture and the embedded emissions in your bike, to the extent those are oil/gas/coal based.
[1] https://ecf.com/news-and-events/news/how-much-co2-does-cycli...
Comparisons are frequently made in other domains as well -- e.g., cyclists measure power outputs in watts.
The rule of thumb I've understood is that the brain runs on about 20 watts (peak power), and a human on average runs on about 100 watts. That's 8.64 MJ/day, which converts to 2064 food calories (kcal), at 4184 J/kcal.
If you wanted to compare them, cite the percent that the CPU uses of its PSU, or cite the percent that human brains use of global energy consumption. Otherwise they're completely different things that have no meaning in relation to each other.
The comparison is valid if you accept the analogy, although I think you could easily dispute the analogy. E.g. you could just as easily view the network of humans in a society as its "brain" and all the things we build (including computers) as its "body".
You could far more easily view the brains of humans in a society as its "brains."
If forced to draw a comparison, you could maybe talk about the energy dissipation of a single neuron and compare that to, say, the energy dissipation of a single transistor. And in that case, considering a human brain is powered by something on the order of 10^10 neurons, the energy efficiency of the brain comes out looking pretty good.
Even this isn't quite true, since it's not fungible, except in the narrow use-case of creating heat[1].
> Whether the energy is carried by glucose/ATP or electric current is just an implementation detail.
This implementation detail means that converting between these forms incurs a tremendous cost (in lost energy) and, perhaps more importantly, is nowhere near instantaneous.
[1] Even the slightly broader use case of heating (e.g. increasing the temperature of a home) can be done with a heat pump, and those don't tend to run off of fungible energy.
When reading the title I thought "They do?"
My laptop consumes the equivalent of approximately a single 60w incandescent lightbulb that used to be required just to write on a pad of paper in the dark.
It's capable of quite a bit more than that, however.
EDIT: Updated power specifications to exact amounts.
2 years, 2 generations before to Sandy Bridge: Core 2 Quad Q9000 (Q1 2009) -> i7 2820QM (Q1 2011) = +100%
Sandy Bridge to 6 generations later, 6.5 years later: i7 2820QM (Q1 2011) -> i7 8550U (Q3 2017) = +37%
Yes, I'm mixing tick-tock cycles or low voltage with "standard versions". So let's see a more appropriate competitor:
Sandy Bridge to 5 generations later, 6 years later: i7 2820QM (Q1 2011) -> i7 7820HQ (Q1 2017) = +50 %
So we had processors that doubled the performance in 2 years (and more than double if we disregard not-so-common Core Quads and take often only two-core mobile processors). Yet, 6 years later, there was only a 50 % increase. Until Ryzen came recently and made Intel to quickly come with Skylake-X, i9s and FINALLY, OH FINALLY, 6 (powerful) cores in laptops, there was no real performance reasons to upgrade, only for heat or thickness, those went down, performance didn't went up so much.
So, around that 2011, those were that crazy times, when you could have:
45 W or 55 W CPU (i7 2820QM or 2920XM)
55 W GPU (Quadro 2000m)
two SATA drives
four DIMM modules
In a non-gaming, back then "standard" size laptop (ThinkPad W520). And under heavy load, anything under Tjunction (say, 95 °C) was good enough. Some laptops with only a single 55 W dissipating fan (such as the ThinkPad W520) often throttled under full CPU and GPU load and couldn't run Turbo with full GPU load, others (such as Dell Precision M6600) had two separate fans and some even came with Quadro 4000m GPU with a TDP of... 100 watts!
So, back then, those laptops ordinarily came with 170+ W power supplies (even over 200 W) and just a CPU+GPU could draw well over 100 W (add the rest of the systems, hard drive here, hard drive there, all four DIMM slots populated, charging the battery) and it would easily go over the twice of that 10 year old laptop with 85 W charger.
And these are workstation models. If you mention gaming laptops, those sometimes had desktop CPUs in them and required 300 W PSUs. Nowadays they also use desktop GPUs and require TWO 300 W PSUs...
Please, don't base observations on 10-year old Tick (probably Penryn?) with 85 W chargers, when 7-8 years ago, a Tock (Sandy Bridge) came with literal concrete bricks as power supplies and really pulled 200 watts. Thankfully, things got dialed back after that, but at the cost of performance until now and current mobile 6-core i9s and Xeons.
And to the lap comments - you really didn't want to have those laptops in your lap when they were plugged in and turbo-ing...
It's more an indication of what your laptop would draw if it were doing GPU-accelerated deep neural net training while recharging a completely drained battery, plus a healthy margin of error just to be on the safe side.
I have a Thinkpad that has an i7-7700HQ, that's a 35W TDP beast and even under very heavy load I still get 90 minutes out of a 72 watt-hour battery and it gets pretty warm (and the TP has excellent cooling), 200W would melt your legs.
And power consumption tends to add up from sources that are otherwise not very power hungry themselves (compared to a CPU or GPU). A few W with the LCD, a few with the SSD/HDD, etc.
My primary laptop has its CPU undervolted (offset) by whopping 170mv, effectively around 30% lower power consumption.
A lot of laptops do actually get too hot (and are too heavy) to effectively be used on laps. Some even require two concurrent power connections from two separate power bricks in order to charge!
My old dusty i7 laptop did get much much too hot though.
These guys measure idle power consumption for all laptops
https://www.notebookcheck.net/HP-Elite-x2-1013-G3-i5-8350U-S...
I have to keep reminding older concerned family members that a light might consume a dollar's worth of electricity per year, or that the bright overhead floods actually consume half the power of the 'low-energy' under-cabinet system (incandescent).
When you talk about "computers using so much energy", the span of inefficiencies could include:
A) Why so much energy per primitive operation (e.g. write bit to data store, XOR two registers)?
B) Why so many primitive operations per desired computation task (sum these three numbers, find the largest of these five numbers)?
C) Why so many computation tasks per unit of economic value delivered?
The article is talking about A). I'm pretty sure we've improved on A) over the history of computation.
I think we've gotten worse on B) per Metcalfe's law (Intel giveth and Gates taketh away).
[1] Called "reversible" because thermodynamics says the only reason you need to dissipate heat/usable energy is because you perform an irreversible operation, and to avoid that you'd have to do your computations using only reversible logical operations -- where the input space has a one-to-one mapping to the output space.
Also what percentage of that 5% goes to mining bitcoin, or re-downloading jQuery for the 5000th time?
A brain uses about 20W. A modern computer has similar consumption, but is many orders of magnitude less powerful.
And for much of those 70 years, energy efficiency was not the top priority.
Our brains are highly specialized, you’re comparing an ASIC strapped to a Z80 with modern general purpose CPU’s.
Our eyes do not see. Our brains do. And the upscalimg from limited data is amazing!
https://www.scientificamerican.com/article/thinking-hard-cal...
> So a typical adult human brain runs on around 12 watts—a fifth of the power required by a standard 60 watt lightbulb. Compared with most other organs, the brain is greedy; pitted against man-made electronics, it is astoundingly efficient. IBM's Watson, the supercomputer that defeated Jeopardy! champions, depends on ninety IBM Power 750 servers, each of which requires around one thousand watts.
https://bgr.com/2016/02/27/power-of-the-human-brain-vs-super...
So highly parallel system are closing in on Teraflops per watt, which sounds amazing to me. But I really have no idea if that's high or low power.
The computing done is value added to energy-use that would otherwise dissipate through the walls and ceiling. Those of us in northern latitudes should be lucky enough to have a couple of servers in the closet helping to pay for heating.
Heating with natural gas or heat pump is more efficient (in terms of BTU per dollar) than using the waste heat from computing (which is resistance).
Heat pumps are not great in cold climates. They are most effective at saving energy when in the heating mode. In a cold climate, however, your house needs more heat as the temperature outside goes down—but the heat pump works less efficiently at lower outdoor temperatures. Below a temperature known as the “balance point,” normally from between 30 and 45 degrees F., supplementary heat is required—and that means expensive electrical heating kicks in. https://www.hometips.com/buying-guides/heat-pumps.html
My home town seldom reached 30 degrees from November to April.
Lots of so-called hyper-heat or low-temp heat pumps (often mini-splits) in cold New England. Affording a heat pump is not a high hurdle. You can get a single zone mini-split (heat/cool) for under $1500 with an install typically well under $1000 without a lot of shopping in most of the US.
Organic brains are also learning and doing language/image/sound processing that far exceed society's computer capabilities.
"Reversing Entropy with Maxwell's Demon | Space Time"
https://www.youtube.com/watch?v=KR23aMjIHIY
Turns out even in case of an "ideal Maxwell'S Demon", the memory operation required to handling pass/stop necessitates certain energy use.
Ads.
1: People will buy what's new and shiney.
2: companies produces what can integrate into what sells
3: loop
I think they meant to say decrease in the parentheses. But that's tantalizing. I'd like to hear more about that.
Old school mainframe power suplies are probably even less efficient per what.
Power consumption is between square and m^n to inter-core bandwidth and count.
So, a high-end smartphone CPU might do 1/5th of the throughput of a desktop CPU using 1/25th of the power.
Just guessing here, by pulling apart what little is in there:
* Total power usage has a big fixed parts, e.g. screen power usage going up linear with brightness and screen size. Or RAM power usage goes up linearly per bit just to keep the memory intact. So Let's assume OP talks about only the 'clocked' components, like CPUs and busses.
* I did hear for single core CPUs how the power usage arose with the square of the CPU speed, unless offset by die shrinking or technological advances. This was one of nature's walls hit by the Pentium IV. Can't find a good reference atm, but here http://www.tomshardware.co.uk/answers/id-3025075/higher-cloc... is a page with the first graph looking square-ish being what happens with power usage when clock goes up on the same machine.
* And throughput is linear with clock speed, if nothing else changes. If you double your clock, you double your throughput just by moving more data over the same bus. So the third line of OP is consistent with this: 1/5 of the throughput is reachable with 1/5th of the clock, thus 1/25 of the power.
* First line might compare 2 ways to raise the throughput: Either rise the speed of 1 core or add more cores. When you rise the speed of 1 core, power goes with the square of this rise. When you add more cores, power goes linear with these cores, assuming your interconnect is free.
* Second line 'between square and m^n' needs some interpreting. It might make sense when for a square n=2, and m measures frequency/throughput/core count/... . I assume he says the real relation might be power usage = throughput^2.3 for example
* When you have m cores, and each is interconnected to its m-1 brothers, you have m*(m-1) interconnect elements using power. This is again a square law. You can get around this by not connecting everything to everything, but this will cost you in data throughput - There will be bottlenecks.
Thats what i can make from it.
https://physics.stackexchange.com/questions/34766/how-does-p...
First answer says P goes linear with f, based on theory.
Second answer goes for cubic or worse based on some deep papers, but a comment says this is dated.
Third answer has a graph from real-world measurements which is clearly worse than linear, and a comment saying there is an exponential part somewhere there too, but it is swamped by linear factors.
WIkipedia, by the way, goes for linear - https://en.wikipedia.org/wiki/CPU_power_dissipation#Sources
I think the most realistic answer I can give is: It is complicated, and OP is a very good nerd sniper ;-)
Phones, on the other hand, are highly specialized - PC's are general purpose. Trade off is often power requirements and performance.
My 8 core 16 thread Ryzen Gaming PC smokes anything my phone can do and i guess if i really wanted to compare i could do a taskper watt calculation and see how they compare.
When i use my phone like a computer, its power draw increases and i get just a couple hours of battery use. play fortnite on your phone and it goes dead quick.
I haven't done a ton of side-by-side comparisons, but it does feel like ARM runs more efficiently per watt.
Phone SoCs specifically address that by having 2+ CPUs of varied power, and having the high-power CPUs completely shut down and powered off when idle.
Same applies to GPU cores that eat significant power when idle, and a lot of power when fully loaded. A PC with a 600W PSU likely features a large GPU that can consume e.g. 300W under full load, doing 60fps photorealistic rendering, or training a neural network. Phones just don't have this kind of hardware (they have much smaller ans less complex GPUs), and, again, manage to keep them powered down when not in use.
Again, a PC drawing 50W or less while doing light office work is a pretty normal thing.
Phones on the other hand are rarely plugged in, and consumers clearly demand more battery life. This has led to resources being poured into cell phone battery capacity and into efficiencies in power usage.
There is simply no comparison to make when we are talking raw compute power. If there were, the DC market would be dominated by x86 processors with stupidly low power consumption. The cost savings from the smaller power/heat footprint would be absurd.
2) Your phone is not nearly as powerful as your laptop despite similar specifications of the CPU speeds. Pretty much everything other than the base CPU clock rate is much, much slower. In addition, your phone CPU down-clocks itself due to thermal limitations where your laptop has a fan and much better heat dissipation capabilities. As a result, it is almost never running at the maximum rated clock speed.
On a YouTube talk I watched (skewering web site speeds and JavaScript overloads), the speaker's iPhone, with CPU specs similar to his laptop, was 25 times slower than his laptop when rendering a (JavaScript heavy) benchmark page.
Where he talks about benchmarked speed vs. marketing numbers: https://youtu.be/4bZvq3nodf4?t=676
Sometimes we stop just a bit too soon. I’ve found a great rule of thumb is that once you’re code complete on a module, put in an extra 10% of work with the profiler to try to improve performance. Keep doing 10% spurts until the performance improvement of the last spurt is less than 5x.
There’s almost always low hanging fruit which will return at least 10x. Sometimes you find multiple places where you get 10x and your code ends up more than 100x faster.
And at that point all your unit tests are written and green so the refactoring is pretty quick. Right? ;-)
It's often the case that optimizing a special case 10 times increases throughput, but makes the user experience worse (because it starts stuttering), makes the code ugly, and allows for timing attacks. So, in many contexts it's actually bad idea.
Half the work in optimization is measuring the correct thing.
I beg to differ.
My Droid Turbo is powered by a 2.7 Ghz quad-core CPU, yet often is incredibly slow. If I want to open Waze and enter and address to navigate to, it's often 10-15 seconds from the time I tap the Waze icon on the launcher until the apps is loaded and the on-screen keyboard has presented itself for typing.
There are significant stutters when I'm trying to scroll while a web page is loading.
Fallout Shelter takes nearly an entire minute to load.
Sometimes, plugging in a USB charger causes the phone to basically become a brick for half a minute or more, as if there are tons of apps deferring processing until I'm on external power.
Overall, it often doesn't feel any more responsive than my original Droid back in 2010, which was only a 550 Mhz single core.
This came up on another thread here from a blog/article titled "software disenfranchisement". I mentioned that I think a lot of it is an unhealthy obsession with writing re-usable code that results in abstractions on top of abstractions piled onto more abstractions to the point where we get the infamous FizzBuzz Enterprise Edition [0] that I'm sure nearly everyone at HN has seen. We also have a problem of people statically linking massive libraries when they only need a small fraction of its code.
> user-hostile decisions that prioritize ads
Definitely a problem in mobile web browsing. Ridiculous that on a data-limited connection, ad networks still try to push video ads or ridiculous JavaScript code.
> But clock rate is truly meaningless as a performance measure when the underlying architecures are different...
Definitely true. ARM gets a lot less done per clock tick than x86. But I imagine a lot of the performance differences are done on purpose as a compromise to reduce power consumption, since ARM is so big on mobile platforms. This is entirely speculation on my part, though.
[0] https://github.com/EnterpriseQualityCoding/FizzBuzzEnterpris...
If you actually use battery intensive apps like google maps' you will quickly realize your battery isn't going to last very long.
[...] There are several factors contributing to the CPU power consumption; they include dynamic power consumption, short-circuit power consumption, and power loss due to transistor leakage currents: [...]
[1] https://en.wikipedia.org/wiki/CPU_power_dissipation
[2] https://physics.stackexchange.com/questions/34766/how-does-p...
As such, future chip will likely be driven to reversible / adiabatic tech because cryptocurrencies, AI and more are guzzling energy ($$$ and running more circuits) and dissipating more heat that can be sensibly handled in a given unit volume (per-socket thermal characteristics; chilled water plant capacity; forced per-cabinet cooling or submersed in mineral oil).
https://www.ncbi.nlm.nih.gov/pubmed/10984064
Also refer - https://www.nature.com/articles/nature13570 by Igor Markov
The future of hardware will be reprogrammable circuits that specialize to repeatable tasks
ICFP 2018 Keynote Address: The Role of Functional Programming and DSLs in Hardware https://www.youtube.com/watch?v=VqYkcGRr8x0
[1]https://www.scribd.com/doc/52657907/Feynman-Lectures-on-Comp...
If you're operating off of a battery, which accounts for a huge amount of consumer computing (especially if phones count), then it starts to matter a lot. Which is why consumer devices tend to be so heavily optimized for efficiency.
If you're operating a datacenter, it also quickly becomes a big part of your costs. Which hasn't necessarily been an immediately pressing concern for Intel, sure, but that may be why companies like Google have been toying with ditching Intel for ARM.
Long ago I had an internship where we were developing a battery powered microprocessor based gadget. We had a prototype with a knob for the clock speed, and we slowed down the clock until the thing started being noticeably slow. Then we chose fixed components for the same clock speed. That's how we optimized battery life.
Note that my thinking is based on relatively simplistic CMOS processors. I don't know how this applies to higher performance computers in cell phones and laptops, with dynamic memory and other goodies.
sure, we pay a high power cost for reliable, deterministic, digital and synchronous computation. but we designed it that way, for that reason. we don't want unreliable or non-deterministic or approximate or whenever computation.
no question that some practices waste power. but we're doing just fine: power per computation (and probably more importantly, for communicating with memory and other systems) is coming down fast. move along, usual moore's law-like behavior, less to see each year...
-Joseph Van Name Ph.D.
Thermodynamics of Computation Wiki https://centre.santafe.edu/thermocomp/Santa_Fe_Institute_Col...
But I’ll take this opportunity to make my usual PSA: any time you are posting about the state of an industry please please date your material
they mean decrease, right?
https://spectrum.ieee.org/semiconductors/devices/the-tunneli...
Woah... really makes you think.
| | nJ/insn |
| MSP430 | 0.9 |
| PIC24 | 2? |
| 1990s StrongARM | 1 |
| LPC1110 | 0.3 |
| Pentium | 10 |
| STM32L0 | 0.23 |
| Ickes DSP 2008 | 0.01 |
| Subliminal 2006 | 0.0026 |
StrongARM: http://www.researchgate.net/profile/Kristofer_Pister/publica...
LPC1110: http://www.nxp.com/documents/data_sheet/LPC111X.pdf
Pentium: http://www.newscientist.com/blog/technology/2006/08/explodin...
Ickes DSP 2008: http://www-mtl.mit.edu/researchgroups/icsystems/pubs/confere...
Subliminal 2006: http://web.eecs.umich.edu/~taustin/papers/VLSI06-sublim.pdf 2009: https://web.eecs.umich.edu/~taustin/papers/TVLSI09-sublimina...
This is sort of comparing apples to oranges. The Pentium (all of them) uses wildly varying amounts of power for different instructions, and has 32-bit or 64-bit instructions, with hardware floating point. The STM32L0, LPC1110, and StrongARM are 32-bit processors with no hardware floating point. The MSP430 is a 16-bit CPU, while the PIC24 is an 8-bit CPU. The Ickes et al. device includes a 16-bit FFT accelerator and only runs at 4MHz, but it was only fabricated as a prototype; you can't buy it. The Zhai et al. Subliminal device, also only fabricated as a prototype, only runs at 833kHz, and it doesn't even include an integer multiply instruction, but its somewhat limited ALU is 32 bits.
However, _all_ of these numbers are far from the Landauer bound (kT ln 2). Suppose that we don't make any concessions to reversibility in our CPU design, like Metronome and its successors. A 32-bit instruction, then, erases the 32-bit register where its result is stored, costing 32 kT ln 2, or in some situations, an average of 16 kT ln 2. Supposing T = 300 K, 32 kT ln(2) ≈ 0.092 attojoules. That's _over seven orders of magnitude_ better than the prototype Subliminal processor mentioned above, and nine orders of magnitude better than the Cortex-M0-based commercial processors mentioned above.
Wolpert has published https://arxiv.org/pdf/1806.04103.pdf (mentioned in the article, but not linked AFAICT) which gives better expressions for the cost of computation than the Landauer bound. I haven't finished reading it, but it looks really interesting.
I do agree with the article's conclusion though: the human brain runs at 20W, on par with a low-power laptop, but arguably the former does much more with that energy.
Did I miss something?