Re: [asa] NASA - Climate Simulation Computer Becomes More Powerful

From: Randy Isaac <>
Date: Thu Sep 10 2009 - 14:33:12 EDT

Yes, you have a good point. Any specific application will probably have a
more relevant benchmark. Linpack is probably chosen because it is the most
widely available and gives a common reference point across many
applications. That's why I said it was a good proxy for a trend. But it
wouldn't be good for specific quantitative performance of any given

Ah yes, the relevance of the topic? I better watch out before I get
reprimanded by the moderators!!

No, it's not directly related to science/faith issues but there are many
indirect implications. One, as Rich noted when he started this, is the
rapidly growing capability to do very complex modelling, continually
improving the quality and credibility. Another, totally different,
implication was brought up by someone a few weeks ago who asked me whether I
thought the rapidly growing computer performance capability would mean that
someday computers would outdo humans in mental capability, whatever that
might mean. I continue to maintain the answer to that is no, though many
transhumanists would argue otherwise.

----- Original Message -----
From: "Iain Strachan" <>
To: "Randy Isaac" <>
Cc: <>
Sent: Thursday, September 10, 2009 1:55 PM
Subject: Re: [asa] NASA - Climate Simulation Computer Becomes More Powerful

I realise this is going a bit off the topic of science/faith and more
towards geek-talk, but for climate simulation, would not a more
appropriate benchmark be the performance on SPARSE systems of matrices
rather than dense matrices as in linpack. I'm thinking of schemes of
equations on a finite difference grid, where the connectivity is
between neighbouring grid points etc.

Sparse systems cannot match the floating point performance of dense
systems because of the necessity of indirection (you need one
instruction to fetch the memory location where the (i,j) value is
stored and one to actually fetch the value). In dense matrices the
values are all contiguous in memory and hence can go much faster.

As an aside I use the Intel Math Kernel library and Vector Math
Library extensively in my work.


On Thu, Sep 10, 2009 at 4:56 PM, Randy Isaac <> wrote:
> Dave,
> It's a Linpack performance benchmark. More details at
> So it is a representative
> performance
> of a dense system of linear equtions. They also give a theoretical peak
> performance but the Linpack measurements is the main criterion.
> Randy
>> What exactly do these numbers represent? Are they the peak number of fp
>> ops per second if all goes just exactly right eg no cache misses... you
>> know
>> the drill? Are they an expected peak number of fp ops/sec under some
>> reasonable assumptions? Lastly does anyone actually measure the number of
>> fp ops/sec that the application programs actually obtain?
>> I thought Rich's report that the Intel gaming vector floating point
>> instructions were being used for climate modeling was interesting. Years
>> ago when I was looking at the IA32 instructions I wondered if using these
>> vector instructions might pay off for numerical computation.
>> Dave W
> To unsubscribe, send a message to with
> "unsubscribe asa" (no quotes) as the body of the message.

Non timeo sed caveo
(")_(") This is a bunny copy him into your signature so he can gain
world domination
To unsubscribe, send a message to with
"unsubscribe asa" (no quotes) as the body of the message.
Received on Thu Sep 10 14:34:13 2009

This archive was generated by hypermail 2.1.8 : Thu Sep 10 2009 - 14:34:13 EDT