Monday, May 03, 2010

Machine Precision of GPU

About seven day, I have been think what must I do for my final task explanation chapter. My aims that to show GPU simulation have as same as CPU simulation. But based on the result I have problem like I said in GPU Rounding Problem article. In this section, I am just to present to you about my observe how the precision of GPU, that is part of my final task observation.

Machine precision is maximum number on computer that if added to 1, for example, not changed that number 1. This statement can be ilustrated like Figure 1 below:

Figure 1. Machine precision.
where e=machine precision

I was make CUDA code to look how the precise of GPU. I use just one thread and in one block with address index x=1 and y=1. As CPU, for single-precision of floating point numbers have precision 0.0000009, more than that number my GPU device cannot be handled. I use GeForce 9400 GT and all the result of my underflow, precise and overflow, can you see in the Table 1 below.
Table 1 is show to us the GPU precision as same as the GPU's underflow numbers. Absolutly this is for single-precision overview. This result equally same what happen in CPU. Both of have same machine precision. Otherwise, because I am just use one thread, one block and one kind of GPU, could be somebody else of you have different result with me, please give me an information then. So I will get a conclusion about it.

Download
CUDA code (main.cu)
data_GPU.txt
logo

Popular Posts