On the new 670 card molecular mechanics is twice faster than was on the cards of five hundredth series. Thanks NVIDIA team.
(This was tested on Abalone program [url=“Molecular modeling soft”]http://www.biomolecular-modeling.com/Abalone/index.html[/url]).
I expect that this is due to a decrease in memory latency, which in turn is associated with an increase in memory frequency. My expectation is that the latency decreased by 2-3 times.
Is this true?
Is the problem almost completely compute-bound?
Or perhaps you have a lot of atomic operations ? According to the whitepaper on GK104 there was a 3x atomic performance improvement, which I guess would be closely related to the latency improvements that you mention.
Yes, there are atomic operations. But I do not think they eat more than 20-30% of the speed.
There must be something else.
Yet, what about the latency. Do I understand correctly that it improved in the 600th cards?
Sorry for my English. What mean the phrase “compute-bound”?
Well if for example your application has a high bandwidth utilization we would consider it to be bandwidth-bound, ie the bandwidth of the card is the bottleneck of your application.
Now if your problem has a very high arithmetic intensity ( ex many addition and multiplications on the same data element => high FLOP / byte ratio ) we might say that it is compute-bound. If this is the case your performance increase would make sense since the GTX670 has roughly 2x the theoretical performance of the GTX570.
Hope I’m making sense :)
Thank you.
No, the problem is not in the long-term calculations. Conversely card mostly unloaded.
It basically waits for a response from the global memory, accessed by the virtually non-regular.
It would be more quantitative if you can use your algorithm in the roofline model.
What is “roofline model”?
Samuel Williams, Andrew Waterman, David Patterson
Roofline: an insightful visual performance model for multicore architectures
Communications of the ACM, Volume 52, Issue 4, April 2009, pp. 65-76