It's in the GNU date manual: http://www.gnu.org/software/coreutils/manual/coreutils.html#Seconds-since-the-Epoch. The same explanation is in the coreutils info pages.
If you precede a number with ‘@’, it represents an internal time stamp
as a count of seconds. The number can contain an internal decimal
point (either ‘.’ or ‘,’); any excess precision not supported by the
internal representation is truncated toward minus infinity. Such a
number cannot be combined with any other date item, as it specifies a
complete time stamp.
Internally, computer times are represented as a count of seconds since
an epoch—a well-defined point of time. On GNU and POSIX systems, the
epoch is 1970-01-01 00:00:00 utc, so ‘@0’ represents this time, ‘@1’
represents 1970-01-01 00:00:01 utc, and so forth. GNU and most other
POSIX-compliant systems support such times as an extension to POSIX,
using negative counts, so that ‘@-1’ represents 1969-12-31 23:59:59
Traditional Unix systems count seconds with 32-bit two's-complement
integers and can represent times from 1901-12-13 20:45:52 through
2038-01-19 03:14:07 utc. More modern systems use 64-bit counts of
seconds with nanosecond subcounts, and can represent all the times in
the known lifetime of the universe to a resolution of 1 nanosecond.
On most hosts, these counts ignore the presence of leap seconds. For
example, on most hosts ‘@915148799’ represents 1998-12-31 23:59:59
utc, ‘@915148800’ represents 1999-01-01 00:00:00 utc, and there is no
way to represent the intervening leap second 1998-12-31 23:59:60 utc.