Neutron sensitivity of integer and floating point operations executed in GPUs

Abstract

Graphics Processing Units are very prone to be corrupted by neutrons. Experimental results obtained irradiating the GPU with high energy neutrons show that the input data type has a strong influence on the neutron-induced error rate of the executed algorithms. Moreover, when operations are performed using floating point data, the probabilities for the mantissa, the exponent or the sign to be corrupted are very different. We investigate the occurrences of errors in the different positions, evaluating the related effects on the result precision. The reported results and the architecture analysis demonstrate that under radiation, whenever possible, one should favor floating point arithmetic, which is both more reliable and potentially easier to protect than the integer one.