cublas cublasZgemm() slower than expected

4
votes

According to nvidia. cublasZgemm is 6x faster than intel MKL.

However, on my PC (i7 2600, Nvidia gtx560, OS:linux 64bit), cublasZgemm is slightly slower than MKL.

I use the numpy.dot() that come with enthought python distribution which links numpy against MKL 10.3.

A matrix multiplication function using cublasZgemm is compiled in a shared library and is called using ctypes in a python script.

When multiply two 1024x1024 complex matrices. numpy.dot() took 84ms. The ctypes function call spent 110ms while the cublasZgemm() part took 97ms.

I wonder why cublassZgemm is not as fast as nvidia stated?

1
pythoncudactypescublas
The GPU you have is much slower in double precision performance than the one which NVIDIA used to produce the benchmark numbers. - talonmies
To expand on @talonmies' answer, Fermi GPUs that are not used in Tesla cards have been artificially limited to double precision speed one quarter of what the chips are capable. I presume that he did not enter his reply as an answer because you didn't mention which type of numbers you are multiplying. - Roger Dahl
@RogerDahl: No, he did - ZGEMM is double precision complex. - talonmies
Thanks for your answer, too bad that they limit the speed of double precision operation. - lucas peng
@RogerDahl: the GTX 560 GPU does not have "artificially" limited fp64 performance. The GF114 chip it is based on has a different SM from the GTX 580 (GF110), and as such does not have hardware to optimize fp64 performance. You may notice that there are no Tesla cards based on GF114 -- GF114 is designed for consumer-class products. - harrism

1 Answers

4
votes

I wonder why cublassZgemm is not as fast as nvidia stated?

The short answer is because you used a much slower GPU to perform the benchmarking of zgemm than was used by NVIDIA to generate their performance figures. Your GTX560 is probably about eight times slower in double precision performance than the Telsa M2090 that was used by NVIDIA in your link.