spec.org's website shows the CFP2006 rates benchmarks for these kinds
of systems. For 2 CPU, 4 core systems, the benchmarks range from
about 30 to 40. For 4 CPU, 8 core systems, the results range from 40
to 60. Of course, these are just benchmarks and not directly
applicable for the same kinds of processing. These caches can help
speed things up dramatically if the requested memory is in the cache,
and that depends on both the way the application is written and the
size of the data. So for example, a blurring application can access
data across columns, rows and slices of voxels. The distance in
memory from one access to the next can vary depending upon exactly
how the program was written, compiler optimization settings and the
size in each direction of the data. If the voxel is outside the
current cache, the speed of the program will be reduced. There are
some benchmarks for AFNI 3dDeconvolve processing, but none of these
include these particular kinds of systems.
http://afni.nimh.nih.gov/afni/doc/misc/afni_speed/
Hope this helps.
Daniel Glen
On Jul 16, 2007, at 5:00 PM, Himachandra Chebrolu wrote:
> We are upgrading our server to a 4 cpu-Xeon - Dell machine and
> I need to make a decision about which cpu's to get. The Xeon's
> now come with L3 cache (4mb, 8mb, or 16mb). The cost difference
> between four 2.6Ghz-4mb L3 and four 3.2Ghz-16mb L3 is 3,0000$.
> I've seen
> many websites that have posted FEEDS scores for older machines, but
> have not
> seen any discussions regarding this issue. Many application will run
> faster with more L3, but many applications won't.
>
> Anyone have an opinion as how much the L3 will affect fsl
> application processing times (we also run spm, afni, analyze,
> ......).
>
> Thanks for any input.
>
> Dave
|