You should use the point estimate of the reliability, not the lower bound or whatsoever. I guess by lb/up you mean the 95% CI for the ICC (I don't have SPSS, so I cannot check myself)? It's unfortunate that we also talk of Cronbach's alpha as a "lower bound for reliability" since this might have confused you.
It should be noted that this formula is not restricted to the use of an estimate of ICC; in fact, you can plug in any "valid" measure of reliability (most of the times, it is Cronbach's alpha that is being used). Apart from the NCME tutorial that I linked to in my comment, you might be interested in this recent article:
Tighe et al. The standard error of
measurement is a more appropriate
measure of quality for postgraduate
medical assessments than is
reliability: an analysis of MRCP(UK)
examinations. BMC Medical
Education 2010, 10:40
Although it might seem to barely address your question at first sight, it has some additional material showing how to compute SEM (here with Cronbach's $\alpha$, but it is straightforward to adapt it with ICC); and, anyway, it's always interesting to look around to see how people use SEM.