| Science Forum Index » Physics - Research Forum » Floating point energy... |
|
Page 1 of 1 |
|
| Author |
Message |
| Dirk Bruere at NeoPax... |
 Posted: Mon Sep 28, 2009 9:54 pm |
|
|
|
Guest
|
For non reversible computing the energy cost of throwing a away a bit is
(IIRC) kTln(2)
Given something like a double precision floating point calculation, what
is the approximate cost in "bits"? I'm trying to work out how energy
efficient the best modern computers are, based on GPU tech ie 10^12
FLOPS for approx 100W
--
Dirk
http://www.transcendence.me.uk/ - Transcendence UK
http://www.theconsensus.org/ - A UK political party
http://www.blogtalkradio.com/onetribe - Occult Talk Show |
|
|
| Back to top |
|
|
|
| Ian Parker... |
| Posted: Tue Sep 29, 2009 3:09 am |
|
|
|
Guest
|
A much better way to look at this is kT. The cost of computing a bit
is nkT. Assuming 77C or about 350K this gives us
k=1.3806503 × 10^-23
n*2*10^-21 joules per bit. If we assume a 3GHz clock this gives us
6n*10^-12 watts per bit stream. A 64 bit machine will in practice have
several things going on at once. Let us give a ballpark figure of 1000
bit streams. This gives us 6n*10-9
n in practice has got to be at least 20 to eliminate the possibility
of chance errors. Finally we get a minimum energy of 10^-7 watts per
core. Figures for the best computers involve about 5w per core.
It is clear that although Moore's law may be nearing its end as far as
sizes of components is concerned, in terms of energy consumption there
is still a very long way to go.
- Ian Parker |
|
|
| Back to top |
|
|
|
| Dirk Bruere at NeoPax... |
| Posted: Tue Sep 29, 2009 9:05 pm |
|
|
|
Guest
|
Ian Parker wrote:
[quote:bb6d3dd2b1]A much better way to look at this is kT. The cost of computing a bit
is nkT. Assuming 77C or about 350K this gives us
k=1.3806503 × 10^-23
n*2*10^-21 joules per bit. If we assume a 3GHz clock this gives us
6n*10^-12 watts per bit stream. A 64 bit machine will in practice have
several things going on at once. Let us give a ballpark figure of 1000
bit streams. This gives us 6n*10-9
n in practice has got to be at least 20 to eliminate the possibility
of chance errors. Finally we get a minimum energy of 10^-7 watts per
core. Figures for the best computers involve about 5w per core.
It is clear that although Moore's law may be nearing its end as far as
sizes of components is concerned, in terms of energy consumption there
is still a very long way to go.
- Ian Parker
[/quote:bb6d3dd2b1]
If we do the same for a Human brain...
10^10 neurons, each with 10^4 axons each flipping a bit at 1kHz we get
10^17 b/s for a dissipation of around 20W. Which implies brains are
within 10^-4 of theoretical limits. Quite impressive.
--
Dirk
http://www.transcendence.me.uk/ - Transcendence UK
http://www.theconsensus.org/ - A UK political party
http://www.blogtalkradio.com/onetribe - Occult Talk Show |
|
|
| Back to top |
|
|
|
|
