Hacker News Reader

Comment by RamRodification | Hacker News Reader

RBC calculates that running a 40MW data center on diesel would require 12,000 liters of diesel fuel per hour

Holy crap

Checking that out:

  You have: 40 MWh
  You want: tonoil
          * 3.4393809
          / 0.29075
  You have: 1 tonoil
  You want: barreloil
          * 6.8419323
          / 0.14615754
  You have: 6.84 barrel
  You want: liter
          * 1087.4731
          / 0.00091956298

So, roughly 1,100 litres of oil per 40 MWh(thermal).

The generator is likely ~30% efficient at converting heat to eletricity, so bump that up to 3,700 litres/hour.

Even allowing for other inefficiencies, 12,000 litre/hr seems high to me. Though as a practical measure it may be correct. It's within a factor of 4 of the actual demand though.

(Calculations using GNU units.)

Aren't you missing x3.4393809 on those ~1100 litres?

Bingo, that's the error, thanks.

Corrected:

  You have: 40 MWh
  You want: tonoil
          * 3.4393809
          / 0.29075
  You have: 3.439 tonoil
  You want: barreloil
          * 23.529405
          / 0.042500012
  You have: 23.529 barrel
  You want: liter
          * 3740.8121
          / 0.00026732164
  You have: 3740/30%
  You want:
          Definition: 12466.667

Or a bit above 12,000 litres. Article was correct.

No need to complicate things with obscure units.

  You have: 40 MW/30% / (10722.2 W*hr/L)
  You want: L/hr
   40 MW/30% / (10722.2 W*hr/L) = 12435.259 L/hr

GNU Units has builtins for some standard energy equivalent units, specifically "tonoil" and "barreloil". "litreoil" doesn't seem to be included by default (though it can be added via a local configuration file).

    literoil = 1 barreloil * liter/barrel

And the calculation gives me ...10.69 kWh/L

The barreloil value is for crude, diesel may have a slightly higher net energy density. But we're good to 3 sig figs.

I could have bypassed the tonoil step and calculated thermal energy in barreloil directly though saving another conversion.

Otherwise I'd have to remember energy density/L of diesel fuel.

I've added the definitions for litre, pound, and kg oil and can now run:

  $ units --terse '40 MWh/30%' literoil
  12470.969

The fun part though, at least for me, is showing how to do conversions with GNU units, and the equivalences which become apparent doing so.

I looked around a bit in "definitions.units" and there are actually energy densities in there for various fuel sources, including diesel.

  You have: 1 diesel
  You want: Wh/L
   1 diesel = 10103.465 Wh/L

Slightly different value than in Wikipedia, but close enough.

*blink*

TIL, there are!

Thanks.

(For the curious, the definitions file is ~7500 lines, there's a lot in there. GNU version. For those on MacOS, the stock BSD units offers far less.)

I think that's wrong. One liter of diesel contains about 10 kWh of energy [1], so you need 4000 liters at 100% efficiency to generate 40 MWh.

12k liters per hour for the datacenter sounds about right, after accounting for diesel engine efficiency.

[1] https://www.sustainabilityexchange.ac.uk/files/cambridge_reg...

Ramrodification caught my missing tonsoil -> barrels conversion.

https://news.ycombinator.com/item?id=31258490

If we take an estimate of 20GW from https://ccaf.io/cbeci/index (huge variance!), and multiply that up, we can estimate that bitcoin is consuming the equivalent of six million liters of diesel every hour.

(It's not going to be diesel, it's mostly a mix of coal and natgas depending on where it happens, but that gives you an idea of the size)

The Emma Mærsk is powered by a Wärtsilä-Sulzer 14RTFLEX96-C engine, the world's largest single diesel unit, weighing 2,300 tonnes and capable of 81 MW (109,000 hp) when burning 14,000 litres (3,600 US gal) of heavy fuel oil per hour.

So bitcoin is equivalent to 425 Emma Mærsks going full bore continually.

That'd be 6.3 million TEU - which interesting means that for those big ships, a liter of fuel moves one TEU for one hour.

> a liter of fuel moves one TEU for one hour.

That's a great little number, and actually feels very efficient!

It’s the second most efficient method of shipping behind barge (and technically sailboat).

Trains come in third, but they use more fuel.

Don't forget about hybrid freighters :)

https://en.wikipedia.org/wiki/Rotor_ship#/media/File:Norsepo...

TUE = Twenty-foot equivalent unit - exact unit of measurement derived from the dimensions of a 20 ft. shipping container.

So even if it had 150,000 humans on board, the humans might consume less liquids than the engine...

Those humans would also only use 1/6th as much energy (when assuming around 9.2m Calorie per liter of diesel). So the energy used by that engine could also "run" a million humans. Whatever that ship is carrying, it might be more energy efficient to let people carry it...

Burning through four liters of diesel a second/a bathtub a minute is insane.

I'm a bit tired, but I hope I got that math about right.

Globally, that mix is 36% coal, 23% natural gas, 15% hydro, 10% nuclear, 5% wind, and a smattering of other sources [1].

Also note that bitcoin miners that have a choice of where to locate choose regions where electricity is cheap, such as those serviced by hydro plants..."most" is a slight stretch, "about half" would probably be more correct - but yeah, it's not good.

[1]: https://ourworldindata.org/electricity-mix

If only that cheap energy would have been available for yandex, maybe they'd have made a different choice than diesel...

I've also heard machine learning training is done in places with cheap energy because the training isn't usually latency-sensitive (querying the trained model might be). Not all AI magic is equally useful, but translation engines are the example I heard of and that's definitely helping me a ton (works better than google translate) as a foreigner in Germany.

I hope figures like this make developers think twice about about throwing yet another layer of abstraction on to the pile.

How many VMs does an average business actually require, assuming the software wasn't written like total ass?

I feel like not running a datacenter on backup diesel generators is going to have a slightly bigger impact than cutting back a few VMs

How so? Large diesel generators are respectably efficient at perhaps 40%. Not as efficient as a CCGT plant at 60% but also not in a different order of magnitude.

Because a 20% drop in efficiency is substantial, and because the alternative to diesel generators isn’t just CCGT. You could also use solar, wind, nuclear, tidal etc etc

Carbon output of a diesel generator is horrific compared to just about any form of large scale energy production.

To be fair, I've seen devs use 1000s of times, even millions of times the resources required.

It usually all goes wrong the moment someone is granted permission to use a 2nd computer to solve some problem.

If you're allowed to use a second computer, you need at least 4 to match the performance you previously had on the single machine.

http://www.frankmcsherry.org/assets/COST.pdf

It can't happen unless software buyers were ready to pay for more development time. That abstraction layer and total-ass quality is frequently the only way to meet the deadline and budget.

If people were ready to pay some upfront costs, it would be possible to lower long-term costs, but that's CAPEX vs OPEX decision that is not in the hands of the software developers.

If one is aware of the basics of performance optimization, you can often find low hanging fruit even in Python.

Basic profiling can often find bottlenecks that are easy to fix, usually just a result of a simple oversight.

No one is asking for every piece of software to be rewritten in Assembly whilst taking advantage of every possible algorithmic improvement.

> Basic profiling can often find bottlenecks that are easy to fix, usually just a result of a simple oversight.

But it requires already scarce developers to spend even more time on one task. Those same developers which aready have two missed deadlines in their task queue. What should be done is always nice, but typically doesn't match reality.

How much time is spent in milliseconds crunching data aka doing work, vs sitting around waiting to work?

CPUs use much less power while idle - if there are too many idle CPUs, that's either because the datacenter is not efficiently managed (e.g. machines are always on regardless of load), or because the standby capacity is needed at short notice.

Which sounds like something a search engine/big company like Yandex would need, this rendering performance a little less relevant since the CPUs aren't pegged at 100% all the time

No, but you see, we need a distributed Kubernetes cluster because that would totallypad out my resume, uh, I mean, make development easier!

And dynamically typed languages.

One of the Finnish news stories stated that as of now, the municipality is receiving only one fifth of the normal amount of waste heat they use for district heating. That sounds like they have shut a lot of things down to cope with the loss of power.

Estimating 10W per core at full load, 4M CPU cores can get a lot done too.

It's like a medium sized container ship.

Reply on news.ycombinator.com