CPU advice

[Dave Nunez]

Hey Guys, I'm going to be getting a new workstation soon and was
thinking about an Intel Core i7-3930K @ 3.20GHz or a single (for now)
AMD Opteron 6274. Any advice if I should go with a higher core
count(amd) or less cores for more speed (intel) ? one thing that
bothered me about the opteron is that the FPU is shared by two IPU's.
I'm planning to make the system my main Eureqa workstation.

Cheers, -d
--
When things get too complicated, it sometimes makes sense to stop and
wonder: Have I asked the right question?
-Enrico Bombieri

[aidabet]

Hello Dave,
The 3930 is ranked the no2 chip, bettered only by the 3960.
http://www.cpubenchmark.net/cpu_lookup.php?cpu=AMD+Opteron+6274
The AMD Opteron 6274 is right up there in 6th place, and with 16 cores
as opposed to Intel's 6 cores, it's a tough call.
The I7 Sandy Bridge and Ivy Bridge chips are the coolest running
performance chips on the market, and since you're going to be working
the machine hard with Eureqa, I am leaning toward cooler running, that
having more cores.
Nevi

DN> Hey Guys, I'm going to be getting a new workstation soon and was
DN> thinking about an Intel Core i7-3930K @ 3.20GHz or a single (for now)
DN> AMD Opteron 6274. Any advice if I should go with a higher core
DN> count(amd) or less cores for more speed (intel) ? one thing that
DN> bothered me about the opteron is that the FPU is shared by two IPU's.
DN> I'm planning to make the system my main Eureqa workstation.
DN> Cheers, -d

[engr.student]

This leads to an interesting question: how do we characterize the
performance of a particular system at running Eureqa/formulize?

There are good questions to think about:
- what are the unique breeds of tasks Eureqa runs?
- how do we measure the performance of a system at those tasks?
- how do the system components (memory, board, bus, OS, chip, cores,
temperature, etc...) relate to the performance of Eureqa at its
tasks?

Note: one of the things that drives me crazy is the inability of
computer guys to use the Buckingham Pi theorem. They like to use
purely linear fits - y = a1*x1+a2*x2... but they do not like
combinations and interactions like y = a1*x1+a2*x2+a12*x1*x2 ...
higher order combined terms. If you are going to build a benchmark,
build one that is capable of handling higher order polynomial, power,
and exponential terms and interactions. heh. I have had to deal with
measurement systems that are incomplete - Buckingham Pi would say that
if you are going to describe a rectangle without using sides - say by
using perimeter, then you need a second metric that is correlated to
the first - area. I'm sure that inability in creating algorithms
extends to characterizing algorithms or relating that performance to
computer hardware. </rant>

These are simple to ask, but profoundly tough questions. How do you
make a "basis" of tasks for Eureqa? A clearly engineered answer to
that question can allow clear statements like "this system is
unilaterally better than that one". It can also answer things like
"improvement in this one area of code will give the best improvement
in the software". Those sorts of statements have economic value.
They sell computers. They sell software. They inform development of
other connected technologies, machine learning, cloud services.

This is what I get for being up so late at night...

On Apr 14, 12:04 am, aidabet <aida...@cox.net> wrote:
> Hello Dave,

> The 3930 is ranked the no2 chip, bettered only by the 3960.http://www.cpubenchmark.net/cpu_lookup.php?cpu=AMD+Opteron+6274

[Dave Nunez]

Hey Nevi, I was just thinking about the same thing, will go with the
intel chip (which from my understanding has better floating point
performance over amd's among other things) Eventually will start a
private network as well.

engr.student:

Eureqa being a highly parallel software is meant to be run on clouds
and private networks..the more cores one has tackling a particular
problem, the faster the solution can be explored. I'm still not
really sure how it's working under the hood, ie: do different building
blocks get assigned to different cores (along with their populations)?
or does the entire problem get explored sequentially among all the
cores available ? is the population size generated based on the core
count available ? maybe Michael can enlighten us with this one.

Cheers, -d

> --
> Eureqa Formulize ( http://www.nutonian.com )
> -------------------------------------------------
> Unsubscribe: eureqa-group...@googlegroups.com
> View Group: http://groups.google.com/group/eureqa-group

[Michael Schmidt]

The more cores the better. Each core allows testing out more candidate models at once. So doubling the number of cores literally doubles the performance.

Note: some cpus now also list the "hyper-threaded" cores or similar, but these extra cores don't add much performance, the number of full cores is what's important.

I highly recommend the i7 i7-3930K, I haven't tried the recent AMD chips though.

Michael

[Dave Nunez]

got the i7-3930K last week, it's really fast. -d