G70 float precision issues

For a GPGPU application, I’m summing up some floating point numbers (say, in the order of 500 floats) to a render target. What concerns me: On pre-G70 NVIDIA cards (FX4400, FX3400, GF 6800), the results are exactly the same. On G70-based cards (7800 GTX, FX 4500) results vary compared to the pre-G70 results in the order of about a tenthousandth part. How can this be? What does the G70 do differently? Which results are to be considered more “correct”?

Thanks for any insights.

It can be anything, from different precision, rounding, instruction ordering, etc. It’s a different hardware after all.
You might want to ask GPGPU specialists on http://www.gpgpu.org/forums/

This topic was automatically closed 183 days after the last reply. New replies are no longer allowed.