the x86-64 has twice as many general purpose registers and twice as many sse registers as the x86 which should result in less cache/memory accesses.
it's 80 when using fpu, and 128 when using SSEx set.
So no, amd64 don't use 128 except when told to do so (mfpmath=sse will higher it as it will use SSE instead of FPU)
And it have nothing to do with amd64, x86 also use SSE@128bits
The Pentium 3 and Athlon XP have SSE, the Pentium 4 and Athlon-64 have SSE2. Also, all newer-to-those chips have SSE(2).
SSE uses 32 bit floating point numbers, or (SSE2) 64 bit double precision floating point numbers. The reason why the registers are 128 bit long is because SSE works with vectors of 4 floats / 2 doubles instead of individual values.