[hpsdr] Intel's benchmark: GPU vs. CPU

[Charles Brain]

All though this is totally unscientific and I have only just got it going.

I replaced the 16 bit MMX FFT code in my DVB-T
transmitter software (for the USRP2). With the CUFFT library running on
an  NVIDIA GTX 465 GPU, it has made virtually no difference 
to my overall CPU loading. I suspect the overhead of reading and writing 
to the GPU device outweighs the benefits of using it and just doing the FFT
on the board is an unfair comparison. 

But as I said I have only just got it going after wasting most of 
the morning finding that you have to create the plan and call cufft
from the same process thread. Yes I did read something along those 
lines here a few days ago! (I was having one of my many stupid moments).

- Charles G4GUO
  

http://www.g4guo.blogspot.com/
----- Original Message ----- 
From: "Hermann" <hvh.net at gmail.com>
To: <hpsdr at lists.openhpsdr.org>
Sent: Friday, June 25, 2010 12:55 PM
Subject: [hpsdr] Intel's benchmark: GPU vs. CPU


> ***** High Performance Software Defined Radio Discussion List *****
> 
> Hi All,
> 
> an interesting article has been published by some Intel scientists:
> Debunking the 100X GPU vs. CPU myth: an evaluation of throughput
> computing on CPU and GPU.


 1277469124.0