Fast forward 10 years

[Manny]

Well, it's been a while since I've coded in FORTRAN. I have a project
that I want to tackle that will require some number crunching. Now,
I'm in the process of upgrading the guts of my PC but need to stick to
a small budget. It will be used for coding and FEM/FEA apps. AMD
chip's low prices (vs. Intel) are appealing. I remember way back when
(~ 10 years ago) AMD seemed better suited at floating point
computations--I can't remember why. Fast forward 10 years and I find
myself having to google the net for the latest benchmark results and
learning about SSEx implementations, etc... Problem is that most
benchmarks today seem geared towards gaming and video editing. Not
very helpful to me who does neither. So my questions are:

1/ What are the FLOPS benchmarks still used today. So far I've found
Whetstone, Linpack (CPU-world) and Spec.org's CFP2006.

2/ I see that the latest Intel chips way outperform the latest AMD
chips in a wide range of gaming/video benchmarks, but the
aforementioned FLOPS benchmarks seem to narrow the gap. So I'm
assuming that AMD is still somewhat strong in FP computations (CPU
clock for clock that is). Is that your take?

3/ In what way can a L3 cache help improve floating point
computations? Does one need to code differently?

BTW, I hope that this forum is still alive. Too bad the original
FORTRAN group got netjacked.

[Terence]

Yes. One page of spam for every one honest entry. (Sigh) :o(>=

I suggest you look into the use of the nVidia graphics chip and the
free Cuda Fortran compiler that comes with it.
You can then forget about the usual computer chip and its speed as
this graphics chip is thousands of times faster than any commercial
CPU for a home or business computer (it has to be for real time high
speed graphics). The latest news is that the newest chip has doubled
the normal and compleax float point precisions.
Your code then runs on the graphics chip and not the CPU chip.

[e p chandler]

On Oct 27, 5:54 pm, Terence <tbwri...@cantv.net> wrote:
> On Oct 21, 7:55 am, Manny <mgim...@gmail.com> wrote:
>
> > Well, it's been a while since I've coded in FORTRAN. I have a project

> > FORTRAN group got netjacked.
>
> Yes. One page of spam for every one honest entry. (Sigh)  :o(>=

At work where I really need to use a web interface I've given up on
reading CLF via google groups. At home I have set my newsreader to
deal with "eternal-september"'s feed. Fortunately my ISP does not
block the port I need.

> I suggest you look into the use of the nVidia graphics chip and the
> free Cuda Fortran compiler that comes with it.
> You can then forget about the usual computer chip and its speed as
> this graphics chip is thousands of times faster than any commercial
> CPU for a home or business computer (it has to be for real time high
> speed graphics). The latest news is that the newest chip has doubled
> the normal and compleax float point precisions.
> Your code then runs on the graphics chip and not the CPU chip

Most of this is beyond me. I undestand the 8 bit chips where
minimizing the instruction count does most of the job. I don't know
anything about SSE or OMP or MPI or co-arrays or the proper switch
options for g95, let alone dealing with cache, out of order execution
and partial register stalls.......

Optimizing 8088/8086 or '386 .... forget it!

--- e

[Terence]

Maybe I should have said that the graphics card in your computer may
very well be an nVidia unit.
If you are a gamer, architect, or into pattern recognition or signal
processing in any way, or possess an Apple product, you probably have
one of the nVidia chips.
This company is now pursuing the area of high speed processing other
than the gaming, 3D modeling and military areas, and going further
into the utility, aerospace, and university areas and other high-end
mathematical processing requirements, including the use of a Fortran
compiler than runs on their "graphics" chips, a-la the "FPU" concept
of old.