Matrix multiplication ERRORS & few thoughts on CUDA Basic programming errors need correction

[MisterAnderson42]

You should have a look at the matrix multiplication example in the CUDA programming guide. As was already pointed out, your kernel has issues.

i get obsure results from this and even after threads synchronisation timing varies a lot , i cant launch more than a 5000x5000 matrix in this

i have a 9800 GX2 , and the cubin file always show 10 register usage . dunno whats the problem

clock() is an extremely imprecise time measurement. Your kernel call is probably completing in ~milliseconds or less. You need a higher resolution timer. I.e. gettimeofday on linux or queryPerformanceCounter on windoze: these are wrapped in a cross platform way by CUT which comes with the CUDA SDK. If you want extremely high precision timing of just the kernel launch, use the timers in the CUDA event API (read the programming guide).

1. how is FSB limitation overcome in cuda , as intel cpu’s bottleneck on multicore performance over 2 cores , as in consecutive memory access from all the threads.

What do you mean? There is no FSB on the GPU. The GTX 280, for instance, has a 512-bit memory bus tied directly from the RAM to the memory manager on the GPU. It is capable of feeding the GPU with ~140 GiB/s of bandwidth (read or write).

2. Maximum sizes of each dimension of a block: 512 x 512 x 64 , Maximum sizes of each dimension of a grid: 65535 x 65535 x 1 . i wanted to know how are these limitations fixed

Because the hardware scheduler was designed that way.

Note that you can’t always run 512 threads in a block depending on the number of registers used in the kernel.