Software analysis led to the conclusion that the limit is caused by the resolution of the
single-precision floating-point format used by a 32-bit device and software.
All the implementations are done on DE5 FPGA board using
single-precision floating-point computation.
The algorithm employs truncation in a
single-precision floating-point environment after converting the binary32 format to a denormalized binary fraction.
Today's FPGAs incorporate hardened digital signal processing (DSP) blocks are capable of implementing efficient 27-bit x 27-bit multipliers that provide extra bits above the normal single-precision 24-bit mantissa requirement for
single-precision floating-point arithmetic.
Assuming peak rates for both systems, you can see that the PS3 is highly unbalanced and favors
single-precision floating-point performance.
In addition, MATLAB 7 offers built-in support for integer and
single-precision floating-point math, as well as language features for managing and analyzing larger data sets.
The resolution will not, however, affect how the data is stored because the
single-precision floating-point data type will accommodate either 12- or 16-bit data.
--added
single-precision floating-point instructions
Except for IEEE denormals, this allows
single-precision floating-point values to be used directly in double-precision arithmetic and branches.