IEEE-754 standard on the NVIDIA GPU (sm_13)

Question

If I perform a floating-point operation (single precision) on the host and device (GPU arch sm_13), will the values be different?

+3

Abhinav Apr 26 '12 at 13:22

1 answer

Jonathan Dursi · Accepted Answer · 2012-04-26T14:01:55+0000

A good discussion of this can be found in the NVIDIA white paper . Primarily:

IEEE-754 is being implemented in almost everything at present;
Even between the correct implementation of this standard, you can still see differences in the results (cool, Intel does an 80-bit internal for double precision), or high optimization parameters with your compiler can change the results.
2.0 NVIDIA IEEE-754 , ,
- - ,
- , ,
- CUDA () , , , , ( ), , , ieee-754.
1.3 ieee-754, , , ; ( - , - , FMA, ).
1.2 , ieee-754, .