Pinned topic Performance problem with PS3
I am currently using a PS3 to do a series of matrix inversions using the CELL LAPACK routines.
See the attached file where I am inverting a 1000x1000 DP matrix using DPOTRF and DPOTRI which takes 2 seconds to complete using 6 SPES (LAPACK_NUMSPES=6) as opposed to a runtime of 0.8 seconds in my Pentium 4 3.4Ghz. When I run the command 'spu-top' I see the 6 SPUs at 100% while the program is running.
I used the Makefile provided with the CELLSDK3.1 and this program compiles without any warnings nor errors.
I tried both SP and DP cases and the Pentium 4 running on 1 thread beats the CELL for any matrix size.
Why is the performance so low??
Thank you in advance
NLai 27000238R315 Posts
Re: Performance problem with PS32014-07-01T07:32:38ZThis is the accepted answer. This is the accepted answer.
I wonder you might need to take advantages of Cell's 128 of 128-bit vector registers and reduce the use of LOOP statements in favor of parallel programming.
I have all modern Intel and AMD built, no one comes close to the Cell in my PS3 core-to-core; Intel is OK, but the 6-cores AMD is not only poor in a single core performance, but also running very very hot like a heater, better to run it in winter.
I would suggest IBM make another round of PC, this time call it SPC for Super PC using a version of the Cell processor to target university professors and learners, since they would have never found the Cell processor "difficult to program" like what Sony did,