What steps will reproduce the problem? Compile the source code with latest cuda driver and sdk 3.2. Run it with a tesla c2050/2070

What is the expected output? What do you see instead? For adult dataset, the accuracy is 82.67% on tesla c1060 with cuda 2.3. However the accuracy drops to 48.20 on tesla c2050/2070.

What version of the product are you using? On what operating system? CUDA 3.2 Fedora Core 13

Please provide any additional information below.

Nothing is modified in the source code. But the number of support vectors as well as the bias value are always not the same on c1060 and c2050 cards. Please give me some suggestions on what could go wrong with the new card.

Thanks!