How large are the numerical errors of the signal.iirdesign method from SciPy?

I am trying to implement a digital filter on an FPGA, but at low frequencies, the filter is often unstable. I am not sure if the precision errors come from the generated coefficients or the hardware implementation, although I am leaning towards the former since the bit widths of the internal signals in the FPGA are very large. For generating the coefficients, I mostly used the Chebyshev II method. The filter is a fourth order SOS filter with a sampling rate of roughly 2 kHz.

As an example, I tried a 17 Hz to 180 Hz bandpass filter, which worked, but when I lowered the lower cutoff to 8 Hz, the filter became unstable.