Hello Ross:
thanks a lot for such kind comments. That's really helpful.
best
Albert
On 09/18/2012 06:15 PM, Ross Walker wrote:
> Dear Albert,
>
>
>> I am wondering regardless the performance, which GPU may give us more
>> reliable results for Amber? Tesla or GTX? I just heard from some friends
>> that they prefer Tesla which is also more expensive than GTX. And they
>> said the results from Tesla is more close to CPU performance.
> This is somewhat of a difficult question to answer given you don't really
> define reliability properly. The answer you want, or the one you want to
> advertise, really depends on which department of a company / university /
> lab you work in.
>
> If you work in new age marketing then you care most about the color GREEN.
> In this case what is most important is RELIABILITY per WATT. In which case
> the Tesla card is better.
>
> Now, if you work in accounting and finance then you care more about the
> color RED. In this case the more important metric is RELIABILITY per
> DOLLAR in which case the GTX card easily wins.
>
> If you work for a large government lab then you probably care more about
> the color ORANGE. In this case you believe that ECC is the most important
> metric (to be strictly correct you believe that ECC backed FLOPINESS is
> the most important metric) in which case the correct measure is
> reliability in terms of ECC ERRORS per unit time. In this case the GTX or
> TESLA card wins depending on how you choose to handle divisions by zero in
> your metrics.
>
> If you are a PI at a cash strapped university in California that spends
> indirect money on who knows what and therefore believes it also has the
> right to squeeze the stone of research grants to get even more blood out
> by having recharge for everything including computers in terms of $'s per
> rack unit of space, regardless of what is actually in those rack units,
> then your favorite color by decree is a very specific Navy Blue and you
> care most about RELIABILITY per UNIT of space in which case the Tesla PCI
> modules are the best option.
>
> If you are me, whose favorite color is BLUE, then you probably care more
> about how much science you can get done in which case you care more about
> RELIABILITY per NANOSECOND of simulation in which case the high end GTX
> cards do better.
>
> If instead by your question you simply meant scientific reliability of the
> results, rather than physical reliability of the hardware, then I am sorry
> for overcomplicating things. In this case there is absolutely no
> difference. Both cards do the exact same calculations (actually the exact
> same silicon) and solve the exact same mathematical expressions in the
> exact same way.
>
> Hope this helps.
>
> All the best
> Ross
>
> /\
> \/
> |\oss Walker
>
> ---------------------------------------------------------
> | Assistant Research Professor |
> | San Diego Supercomputer Center |
> | Adjunct Assistant Professor |
> | Dept. of Chemistry and Biochemistry |
> | University of California San Diego |
> | NVIDIA Fellow |
> | http://www.rosswalker.co.uk | http://www.wmd-lab.org |
> | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
> ---------------------------------------------------------
>
> Note: Electronic Mail is not secure, has no guarantee of delivery, may not
> be read every day, and should not be used for urgent or sensitive issues.
>
>
>
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Sep 18 2012 - 10:00:03 PDT
