INTEL XEON vs AMD ATHLON

Don Holmgren djholm at fnal.gov
Wed Jun 12 11:28:02 EDT 2002


On Wed, 12 Jun 2002, Ivan Oleynik wrote:

> I sent email a couple days ago asking for advice re linux cluster vendors
> and critical configuration issues (I am going to summarize my findings
> after getting all the responses). It turned out that the choice of the
> processor is the most critical one. I had an impression from literature
> that Athlon MPX goes head by head with Xeon in terms of CPU speed for
> heavy numerics, but it is considerably less expensive.
> 
> Surprisingly, I received several responses that contradict to my view. In
> particular, Xeon is only 5-10% more expensive, but it is more powerful
> since it is currently running at 2.4 GHz, has 400 MHz front bus plus 512
> Kb of secondary cash as compared to 1.76 GHz, 266 MHz and 384 Kb for
> Athlon. Also, the newest linux kernel will provide multithreading on Xeons
> that could also result in substantial boost for numerical calculations. 
> 
> I think this is old question, but I would like to know what experts
> say: Is it really true that the new Xeon is superior than Athlon as far as
> heavy numerical calculations in cluster environment are concerned?
> 
> Ivan Oleynik

As has been noted several times, it depends upon your application.  For
the clusters in our project, the application (lattice QCD) on a single
node is very sensitive to memory bandwidth, and on a cluster to PCI
performance and memory bandwidth.

I have comparisons posted as follows:

1) http://qcdhome.fnal.gov/qcdstream/new_compare.qcdstream
  This shows math kernel (su3 matrix algebra) performance both in-cache
  and out of cache, C-language (labelled "MILC") and SSE-optimized, on 
  Athon 1600+ and 1800+, and on Xeon 1.7 GHz (Foster, Rambus) and
  Xeon 2.0 GHz (Prestonia, interleaved DDR).  All code built with gcc.
  In-cache without SSE, the Athlons tend to win.  With SSE, or in memory
  with and without SSE, the Xeons win.  The SSE codes are optimized for
  Xeons - I have no idea whether codes specifically written to
  3DNow! would do substantially better.  See
  http://qcdhome.fnal.gov/qcdstream/ for information about this
  benchmark.

2) http://qcdhome.fnal.gov/benchmarks/amd_xeon.html
  This shows comparisons of performance on very small (4 node) clusters
  with Myrinet interconnects of the full lattice QCD code on 1.2 GHz
  Athlon, and 2.0 GHz Xeon (Foster).  Also shown are netpipe and
  gm_allsize comparisons.  Note that both flavors do not have
  particularly good PCI buses - the Athlons here have 33/22 buses, and
  the Xeons have the poor 64/66 buses of the i860 chipset.

3) http://qcdhome.fnal.gov/benchmarks/e7500.html
  Comparisons similar to #2 of E7500-based 2.0 GHz Xeons with i860-based 
  1.7 GHz Xeons.  Shows improvement of PCI bus on E7500 motherboards
  compared with i860 (and AMD MP, though that's not a fair comparison).

We picked E7500-based Xeons for our latest cluster (48 nodes, Myrinet,
operational next month) based on better price/performance on our code.

Don Holmgren
Fermilab



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list