[Beowulf] Electricity cost: a critical survival issue of our ICT infrastructures.
Lux, Jim (337C)
james.p.lux at jpl.nasa.gov
Tue Apr 9 07:01:14 PDT 2013
On 4/9/13 5:09 AM, "Lawrence Stewart" <stewart at serissa.com> wrote:
>I got to thinking, "how bad could electricity costs get?"
>The answer I come to is "not very bad". The potential costs are capped
>by the capital cost of installing your
>own, say, solar generation. If I understand these things correctly,
>which is not likely, it now costs about $1/watt
>for panels, with about a 20% active time, so $5/watt.
>I am not including storage here, so you have to sell 4X your load for the
>20% the sun shines, and buy 1X for the 80% it doesn't.
>The net electricity costs ought to be about 0.
>The green 500 is running over 2 GF/watt these days, or 500 watts/TF. So
>you can buy PV panels for $2500/TF.
>Suppose the installed cost is 4X the panel cost. We're at $10000/TF with
>0 operating costs for electricity.
>Suppose that TF cost $10000 for the computer and AC (I have no idea!).
>Then the capital costs have doubled, but
>the lifetime of the PV plant is probably 15-20 years, vs the 3 year cost
>of the computer.
>I am sure I am way off on this, but if the figures are usable, then for a
>1-2 year hit on Moore's law, you can just buy free power
An interesting trade is whether you would want to site your cluster (and
yourself.. I'm a believer in being able to touch my resources) somewhere
warm and sunny (more solar power, but more AC costs) or somewhere cooler
(less sun, but use outdoor air as a coolant). Of course, somewhere that
it is cool and clear would be ideal.
$3-5/watt is the going price in southern california for a medium sized PV
installation with various incentives taken into account.
"4.90/W for larger commercial systems of 100 kW or more in size.
Utility-sector PV systems larger than 2,000 kW in size averaged $3.40/W
The press release (which has links to some nice reports, as well) makes
the point that non-panel costs are also decreasing. Solar Panel costs are
kind of funny. As of a few years ago, the mfrs could sell all they could
make at the price point they were at, and that in turn was based on the
crummy yield. There wasn't much incentive for improving the yield: some
of the players in the field (Arco, BP) were making solar panels as a hedge
against the future, and for advertising/greenwashing. The flood of panels
from China (with all the currency issues) put a lot of downward pressure
on panel prices (doing in Solyndra, for instance, whose success was
predicated on beating the existing panel prices).
Or here's a more interesting computational question.. Do you scale your
computation up and down as the energy available changes? Stop it when the
sun goes down, and ramp up in the morning as the solar insolation changes.
Clouds reduce your computational rate.
A fascinating optimization question.. Do you buy more panels and/or more
nodes? If you are in the Los Angeles area on the coast, do you have
shorter compute days in early summer, even as the solstice and long days
approach, because June Gloom reduces power output in the morning. Or, do
you install your panels out in the desert and your computers at the coast,
incurring the transport cost for the electricity. (which could be
virtual.. Sell power to Edison at one end and buy it at the other)
More information about the Beowulf