platform for number crunching

[Micha Feigin]

On Thu, 17 Jun 2010 08:38:59 +0300
Orna Agmon Ben-Yehuda <ladypine at gmail.com> wrote:

> On Wed, Jun 16, 2010 at 4:29 PM, Oleg Goldshmidt <pub at goldshmidt.org> wrote:
> 
> > On Wed, Jun 16, 2010 at 1:23 PM, Shimon Panfil <info at industrialphys.com>
> > wrote:
> > >
> > > Hi folks,
> > > I'm looking for affordable workstation for heavy number crunching, not
> >
> > What's "affordable" and what is "heavy number crunching"?
> >
> > For most large-scale scientific/engineering number-crunching physical
> > parallelism (multiple CPUs/cores) is important for performance. Will
> > you benefit from many more than 4 cores? Will anything more than
> > commodity 4 core desktop be prohibitively expensive? Will you benefit
> > / can you afford, e.g., a CUDA-based number-cruncher under your desk?
> >
> >
> Note that while a CPU consumes around 100W at highest P-state (core voltage
> and frequency), based for example on opteron specifications, GPGPUs consume
> around 300W and even more when active. If the current box has a problem with
> dissipating heat, it will not take a GPGPU.
> 
> Furthermore, switching to GPGPU programming has a huge cost in labor. For
> this specific case, in which Shimon is a consultant, my assumption is that
> he works with a different piece of software every time, he will not have a
> chance to enjoy the fruit of his labor.
> 

These numbers are wrong by the way

If you look here for example:
http://www.guru3d.com/article/intel-core-i7-920-and-965-review/18

The difference between a medium CPU at idle (lets push it up a bit to 130W
system draw when Idle) to a core i7 at 100% (lets lower it a bit to 260W at
100%) is over 130W. A low end GPU will take around 30W-40W, a high end second
generation GPU (gtx 285 or tesla c1060, 240 cores, over 600GFLOPS expected,
I'm not talking the peak 1TFLOP) at 100% doing cuda will run at ~160W which
gives you the same delta of ~130W. The card you are talking about is the gtx
480 which has 480 cores and 1.1 TPLOP expected takes 250W, which is a larger
delta but for a pretty nice jump in computing power.

On the other hand, yes, if there are cooling problems now, they are going to
get worse (a whole lot worse) with a GPU, the 285 can go to 80c under load and
tesla to around 65c under load in a ok case with 2 external fans.

On the other hand, if the cpu overheats I would investigate cpu and case issues
first as that shouldn't happen (either the case is not built for the job, the
heatsink is not properly connected or the fans are not up to the job or not
functioning.

As for the software, that is right, with gpgpu you need to write you own
software, but on the otherhand, on the cpu if you are not using sse you are
also under utilizing your cpu for number crunching by a factor of possibly x4-8
depending on the accuracy.

And on that note, GPUs are mainly floating point machines (The new teslas
2050/2070 changed that a bit but still), so if you need double precision or
random memory access or to use existing software/libraries, CPUs are still your
best bet.

>