Power Benchmarks...it ain't really about the chip...
A recent News.com article
http://news.com.com/Chipmakers+admit+Your+power+may+vary/2100-1006_3-6082352.html
there was talk of a power benchmarking metric of performance per watt. This type of benchmarking is not a bad idea. (Frankly it is about time).
However, there are a couple of things about this that should be considered.
The first is that any benchmarking of power is not all about the
processor (single core or multi-core). The processor is only part of
the equation. Today most people think of power consumption and think
only about the processor. It is not.
Does anybody just use the
processor? Not really. After all, what good is a processor without some
RAM to hold the code in? (Answer: Not much.) You can have a great chip, but without an ethernet controller chip and phsyical interface to serve
that data to the (grid, enterprise, internet, etc.) the work that the
chip is doing isn't really of that much use. So when we talk about
power benchmarking, (or any other form of benchmarking for that matter)
it is the entire system that matters. Today, the processor is the main
culprit in heat generation. However, that was not always the case. Back
in the late 80's disk drives used more power and generated more heat
than the CPU ever did. It is not a far reaching leap to think that some
number of years from now when machines have petabytes of memory (no...petabyates is not a typo) the memory system will generate more
heat than the CPU, and if the industry does not look at reducing power
consumption of memory (and other asics as well) we won't really
substantially "fix" the problem. So while Intel and AMD are locked in
combat on the CPU front. I hope the other chip folks are not just in the
stands watching the match.
Also, there is more to life in the real data center world than just
performance per watt of the CPU. Performance per watt per amount of
work done over a certain threshold for the entire system is what is
really important. Why? I am glad you asked. A system has a given set of
work to do over a certain time period. For example: Serve 10,000 web
pages per minute. Let us pretend that vendor A makes a machine that
serves 100 web pages per minute, and the uses 100 watts of CPU power,
and vendor B makes a machine that serves 10,000 web pages per minute
and uses 10,000 watts of CPU power. They are both providing the same
perfromance per watt of the CPU. But, you will need 100 of vendor A machines to do the same work as 1 vendor B machine. However, vendor A will need 99 more network interfaces, needing 99 more interfaces on a
switch (bigger switch to handle the ports means more power used by the
switch). Vendor A needs 99 more disk drives (one per machine or double
that if they are mirrored) All of that "other stuff" uses power and generates heat. However, if
you only look at performance per watt of the CPU you will miss all that
other power and heat.
The good news is that power benchmarking is being developed by more
than just the chip makers and the issues around trying to reduce power
and cooling in the datacenter is being addressed through the Green Grid.
http://www.thegreengrid.org/
Posted at
02:54PM Jun 23, 2006
by rdsesq in Sun |