It limits the buffer size to the respective number. So if you use the fix16 node, the buffers are guaranteed to be never bigger than 16 samples (they might get smaller though if you have a buffer size that is not a multiple of the number).
This can be used to make a consistent modulation update frequency. The cables that you draw between modulation connections are updated once every buffer so if you run the plugin at 512 samples buffer size, these will get updated every 5.8ms which is way too high for certain applications like envelope followers etc.
The obvious solution is to switch to the frame nodes which gets you sample accurate modulation but of the cost of CPU usage, so depending on your use case, the fix## nodes get you a reasonable fast modulation update rate while keeping the CPU usage low (eg. a modulated filter cutoff that is smoothed doesn't need to be updated every sample, so a fix32 block is totally fine for this use case).
