wiki:DGEMM_Performance_RV770

Introduction

I have tested the GPU enabled version of AMCL (version 1.0, avialable at http://developer.amd.com/gpu/acmlgpu/Pages/default.aspx) on HD4780 with Ubuntu 8.04 LTS. It requires gfortran version 4.1.2. In fact, the current version only works with this paticular verison of gfortran. There are two versions of gfortran available on Ubuntu 8.04 LTS. The latest package is "gfortran-4.2" but this does not work with AMCL-GPU. Please install the package "gfortran-4.1" explicitly when you try.

Result

The peak DP performance of HD4870 is 240 Gflops.

Note

According to a paper (http://www.vpac.org/files/ACML_Report_lukep_final.pdf), AMCL-GPU automatically switch between CPU and GPU depending on the size of a matrix. I put Figure 1 from this paper here.

Basically, for N > 750, AMCL-GPU really use GPU but with a smaller matrix, it only use CPU.

Last modified 15 years ago Last modified on Jun 10, 2009 11:54:33 AM

Attachments (2)

Download all attachments as: .zip