energy costs

Josip Loncaric josip at nianet.org
Tue Mar 11 07:40:10 PST 2003


Helms, Scott A wrote:
> [...] I asked about kilowatts per gigaflops [...]

Your concern about KW/Gflop is completely justified.  Both IBM and Cray 
salesmen will happily advertise their power efficiency.  Today's CPUs are 
primarily constrained by their thermal limits, and about half of the time 
spent in CPU design workshops is spent on CPU power management.  As a result, 
we get interesting "features" such as thermal throttling on CPUs, memory, etc.

In theory, one could use reversible physical processes for computing and thus 
recover all of the energy by reversing the computation at the end.  This kind 
of computer would require no net power, although some energy would be needed 
to reach the desired result.  However, this computer would not be allowed to 
forget any of the states it went through.  In some sense, power dissipation is 
the penalty we pay for forgetting -- and thanks to speculative evaluation, 
today's CPUs evaluate then forget even more than is commonly known.

I can't seem to find my KW/Gflop data right now, but I believe that in that 
regard a machine such as Cray X1 compares rather favorably with Beowulf 
clusters.  Both kinds of systems will improve power efficiency each year, but 
their relative merit will probably evolve quite slowly...

Sincerely,
Josip

-- 
Dr. Josip Loncaric, Research Fellow              mailto:josip at nianet.org
National Institute of Aerospace       http://research.nianet.org/~josip/
144 Research Drive                       mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23666, USA         Tel. +1-757-766-1395  Fax +1-757-766-1401




More information about the Beowulf mailing list