C++ binomial tree float versus double preformance

[maciej]

Hey guys, I'm implementing a binomial tree version in C++ to benchmark versus CUDA and I found an interesting issue: changing the CPU code to use doubles(64 bit) for everything instead of float(32 bit) actually increases performance of the CPU code significantly (~3.5x faster.) Anyone know whats going on here, I always assumed floats were faster than doubles???

Also the cpu is a 64 bit processor and the speed improvement occurred when compiled in both MS express edition and g++.

 

[Eugene Krel]

your system could be converting floats to doubles to do the underlying operations.

 

[satyag]

It is probably hardware dependent. Which GPU are you using ?

 

[Chris Matei]

Check this, it might have some valid explanation:

http://www.gamedev.net/community/forums/topic.asp?topic_id=506100

 

[maciej]

Check this, it might have some valid explanation:

http://www.gamedev.net/community/forums/topic.asp?topic_id=506100

Thanks for the link, this seems to be my case also, everything is converted internally to a double. And satyag I'm currently using a g210m for development and a Tesla for benchmarks.