Re: AMD dual core vs Intel core 2 quad
On Wed, 2008-02-06 at 08:35 -0800, Francesco Pietra wrote:
> In the last few days I asked advice about upgrading from a 2-way Tyan
> motherboard Thunder K8WES2895 with two series-2xx AMD dual core opteron to a
> 4-way motherboard. I learned from both Lennart Sorensen and Daniel Tryba that I
> can't use my series 2xx and should buy series 8xx. The price for such system is
> largely beyond what I can afford for private use.
> I am wondering about abandoning totally my present system of dual opteron,
> saving the disks, power supply, and, most important, the 24 GB of Kingston
> KRV400D4R3A (184 pin) memories.
Core 2 is a CONSUMER platform, and I'm not aware of any board which will
take such a large volume of memory - especially obsolete DDR1.
> Abandoning to change to Intel Core 2 Quad Q6600 or 6800. The price is
> attractive and the choice of motherboard is wider, possibly up to a 8-way
> system (hopefully the 184-pon slots can be used). How does such Intel core 2
> compare with dual core opteron in terms of number of independent cpus? To be
> specific: when I run molecular dynamics (MD) on Debian amd64 with my shared
> memory Thunder with 2 AMD dual core, parallelization support OpenMPI
> (everything compiled with Intel ifort/icc), the MD program sees 4 cpus and "top
> -i" indicates four cpu at work. The gain with respect to running MD in serial
> mode is about threefold, i.e. the parallelization is not bad at all. What can I
> expect - from this viewpoint - with four Intel Core 2 quad?
Core 2 is a CONSUMER platform, and there are no multi-socket options.
Xeon platforms based on the same architecture as Core 2 are available,
at a cost. Each quad core chip means 4 CPUs as far as your OS is
concerned, and these are "real" - but the achilles heel of the Intel
platform is memory contention, and for heavily memory-bound codes,
scalability will be poor. Conversely, for CPU-bound codes, performance
will be SIGNIFICANTLY better. Roughly speaking, using a 4-way GROMACs
job as an example, the same job could complete about 25% faster, clock
for clock, on Intel.
> (Compilation of the Forthran MD program with Intel gives a tremendous burst
> with respect to GNU Fortran compilers)
That much is true, but use of a Portland compiler would help on the AMD
platform (probably not by as much, though).