Floating Point DSP¶

Symetrix DSP makes use of Analog Devices SHARC DSPs, which use 32- and 40-bit floating point processing. Compared to 24-bit fixed point processing used in other devices. Floating point processing offers significant benefits to the user in both audio quality and ease of use.

Limitations of fixed point processing¶

A problem with fixed point processing is that if significant gain changes occur, data loss, or worse yet, clipping distortion can occur. For example, consider the case of processing 24-bit audio signal using 24-bit fixed point processing. If in some process, you attenuated the signal by 42dB, the new signal would have only 17 bits of information in it. Seven bits have been lost forever due to the gain reduction. Worse yet is the problem of clipping distortion. If a significant gain is applied to a signal that is already near 0dBFS, the signal will be clipped at 0dBFS resulting in audible distortion. Even if another process downstream reduces the signal level back to what would have been below 0dBFS, the clipping has already occurred and the distortion remains. Fixed-point processes may try to build in some headroom above 0dBFS, but in doing so they must inevitably give up some bits to do so. For example, if 12dB (2 bits) of headroom was built in, a 24-bit system now is only really a 22-bit system.

How floating point processing works¶

In contrast, with floating point processing, all available bits are always allocated to the signal regardless of the signal level. Basically, floating point uses a few bits as an exponent to set the "rough" signal level and then allocates the rest of the bits to storing the actual signal value independent of level. Hence, regardless of the signal level, from -200dB and below to 100s of dB above 0dBFS, the signal is stored with optimum precision and no clipping distortion. The SHARC offers both 32-bit and 40-bit extended precision processing. With 32-bit processing there are 25 bits allocated to storing the signal regardless of level. This means that the precision is always better than 24-bit fixed point by at lease one bit and at low signal levels, by significantly more. With 40-bit extended precision processing, 33 bits are available to store the signal. In Symetrix, 40-bit processing is used within DSP modules, and the data is converted to 32-bit when routed between modules or over the SymLink bus.

Practical benefits of Symetrix's floating point processing¶

What practical benefit does using floating point processing give the user? Gain staging issues between multiple processing modules can be ignored. If one module reduces the signal level by 50dB and another process later on boosts it back up, no data loss will occur. With fixed-point systems, the user must be careful with signal levels to avoid clipping and excessive data loss at every processing module output. The one place where the user must watch the signal level is just before it is sent to the DAC, since all DACs are fixed point. In the Symetrix system, if you noticed that your signal was clipping just before it was being sent to the DAC, you could turn it down right at the output to fix the problem. With fixed point systems, you would need to search back through every processing module to find the source of the clipping.