I don't have any specific SGI data, but I have some GCL vs ACL
data on what I believe is non-numerical large scale munching
(namely Macsyma polynomial stuff).  This is code that was
tuned for GCL, with declarations that ACL would probably not
need.  ACL has  a substantial overall advantage in speed:  

factor of 7  when GCL spends too much time in garbage collection
(several benchmarks)
factor of 4 more typically
On some parts there is a factor of 1-2 if there is no GC, or
if one excludes GC time.

GCL was never faster in my tests.