Floating point DSP, what does floating point DSP mean?

Usually the data width of floating point DSP is 32 bits, which uses 24 bits to form the mantissa and 8 bits to form the exponent. In practical applications, the mantissa defines the precision and the exponent represents the dynamic range.

Fixed-point arithmetic DSP has achieved great success in applications and is still the main body of DSP applications. However, as the requirements for DSP processing speed and accuracy, memory capacity, programming flexibility and convenience continue to increase, since the mid-to-late 1980s, various DSP manufacturers have successively launched their own 32-bit floating point arithmetic DSPs.

Compared with fixed-point arithmetic DSP, floating-point arithmetic DSP has many advantages:

The dynamic range of floating-point arithmetic DSP is much larger than that of fixed-point arithmetic DSP. Every time the word length of fixed-point DSP increases by 1 bit, the dynamic range expands by 6dB. The dynamic range of the 6666-bit word length is 96dB. Programmers must always pay attention to the occurrence of overflow. For example, when doing image processing, if the image is rotated, moved, etc., it is easy to cause overflow. At this time, either continuous shifting and scaling or truncation are performed. The former consumes a lot of program space and execution time, while the latter quickly causes image quality degradation. In short, the performance of the entire system is degraded. Similar problems also occur when processing low signal-to-noise ratio signals, such as speech recognition, radar and sonar signal processing. The dynamic range of the 32-bit floating point operation DSP can reach 666536dB, which not only greatly expands the dynamic range and improves the calculation accuracy, but also greatly saves calculation time and storage space because scaling, shifting and overflow checks are greatly reduced.

Since the floating-point operations of floating-point DSP are implemented in hardware and can be completed in a single cycle, its processing speed is much higher than that of fixed-point DSP. This advantage is particularly prominent when implementing high-precision complex algorithms, providing a guarantee for real-time processing of complex algorithms.

The bus width of 32bit floating-point DSP is much wider than that of fixed-point DSP, so the addressing space is also much larger. On the one hand, this provides the possibility for large and complex algorithms, because the provincial DSP target subroutine has used dozens of MB of memory or more; on the other hand, it also provides the possibility for the application of advanced tool software such as high-level language compilers and DSP operating systems. condition.

The further development of DSP must be the application of multi-processors.