Can you show us your code? Most likely, you're running afoul of floating point precision errors, but that's hard to tell without seeing any of your code. Math::BigInt or Math::BigFloat might help, or maybe tye's Math::BigApprox. These all have different advantages and drawbacks, depending on your speed requirements and precision requirements. Note that it does not make sense to have a (much) higher precision in your result than in your input data.

As the others have pointed out, there are arbitrary-precision libraries out there, but you're going to pay an efficiency price for using them. If you DO use them, make sure you check your results... In the case moritz demonstrated above, -Mbignum gives the correct answer, which you can check by reversing the operation:

$ perl -Mbignum -e'print 3.624920464504015567E+08 + -0.0002095444,"\n"
+'
362492046.4501920123
$ perl -Mbignum -e'print  3.624920464501920123E+08 + 0.0002095444,"\n"
+'
362492046.4504015567
[download]

You can try to circumvent it (with some performance penalty) by using bignum.

$ perl -Mbignum -wle 'print 3.624920464501920123E+08 - 3.6249204645040
+15567E+08'
-0.0002095444
$ perl  -wle 'print 3.624920464501920123E+08 - 3.624920464504015567E+0
+8'
-0.000209510326385498
[download]

Actually I haven't checked which is more accurate, but you can give it a try.

besides bignum, you can also try some multiprecision package. This implies in general having a C library and a Perl module to interface with it. A example is Math::GMP or Math::Pari. Can be better to store everything first as strings and deal with separtely with exponents.

if you want to try Math::Pari, on some systems there is a bug which prevents the build to fetch correctly the Pari lib. If you fetch it yourself and untar it in the module install dir everything should be fine

Double precision floats (which is what 32-bit Perl uses for decimal numbers) have 53-bits of storage for the mantissa, so about 16 digits.

Just like you can't store 257 in an 8-bit integer, you can't store 3.624920464501920123E+08 in a 53-bit floating point. I has nothing to do with floating point numbers really.