lindahl at csb.stanford.edu
Sat Mar 20 00:39:01 CET 2004
Just to confirm what David said - yes, there was a hardware issue with
all recent AMD processors, including Opterons.
I have worked around this in the code starting with 3.2.1, but since
I'm partly under NDA with AMD (which they still haven't released me
from) unfortunately I cannot comment on exactly what the bug is. If I
understand it correct they are still trying to find a software fix.
I cannot prevent you from running a diff on different versions of
x86_sse.s though :-)
Compilers really don't use the parallel SSE instructions at all, they
just work on one element in all cases I've seen. And, apart from
Gromacs the only SSE-tuned programs I've seen are simple things like
video encoding and BLAS that always go through memory in a nice linear
way. Those won't be affected by the it.
I wouldn't worry about going with Opterons because of this - all CPUs
have errata, including Intel ones...
On Mar 19, 2004, at 8:42 PM, David wrote:
> On Thu, 2004-03-18 at 18:49, Thomas Veith wrote:
>> Hi folks,
>> Here we are going to get a cluster which is budgeted for approx. Euro
>> To get most bang for our bucks, we thought of going the
>> AMD-Opteron/AMD64 way; but recently I read here by browsing on your
>> ML that there exists an Opteron Bug, but I couldn't find any more
>> info to this.
>> So my question is: Is there a hardware bug in (SSE/2, anywhere)
>> Opteron, is it fixed, was it related only to gromacs or could this
>> affect any hand-coded assembly?
> Yes there seems to be a hardware bug in all SSE capable Athlons, but
> I am informed correctly gromacs 3.2.1 has a workaround. The bug was in
> an obscure instruction relating to using unaligned memory. It is very
> unlikely that this affects other code, as compilers tend to avoid such
> instructions for performance reasons.
> I think that Opterons are a good choice (I have some myself and posted
> benchmarks last fall).
>> Thanx in advance for your replies :-)
>> Best regards,
>> gmx-users mailing list
>> gmx-users at gromacs.org
>> Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-users-request at gromacs.org.
> David van der Spoel, PhD, Assist. Prof., Molecular Biophysics group,
> Dept. of Cell and Molecular Biology, Uppsala University.
> Husargatan 3, Box 596, 75124 Uppsala, Sweden
> phone: 46 18 471 4205 fax: 46 18 511 755
> spoel at xray.bmc.uu.se spoel at gromacs.org http://xray.bmc.uu.se/~spoel
> gmx-users mailing list
> gmx-users at gromacs.org
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
More information about the gromacs.org_gmx-users