The CX31993 cannot physically output more than 1.0Vrms, so is limited to a little over 30mW per channel at 32Ω.
The 65mW spec must refer to combined output power, but that's usually not how you spec power.
The CS43131 on the other hand can output 2.0Vrms unloaded.
The built-in impedance sensing and auto gain control however will limit the voltage to 1.0Vrms under 32Ω load. Same voltage as the CX31993.
Luckily, this auto gain can be defeated with an adapter like this:
View attachment 257788
It's a simple male to female extension cable, no special resistors or anything.
Plug the extension into your dongle, then your headphones into the extension and the auto gain is defeated.
Using it, you can get a clean 70mW per channel out of the CS43131. That's an increase of 3.5dB or 30% louder than the CX31993.
Here are fresh measurements of the Meizu HiFi and Mblu HiFi to show this:
View attachment 257791 View attachment 257792
As you can see, the CX31993 is
a lot noisier than the CS43131.
The CX31993 also has a distinct increase in noise once the input signal amplitude exceeds -56dBFS. That's the step you can see just below 2mV.
Conveniently, the industry standard -60dBFS test tone that is used to measure Dynamic Range falls just below that threshold, so doesn't trigger the noisier high gain.
Meaning that whatever SNR/DR numbers you see for the CX31993 are fake and don't accurately reflect the DAC's performance under real-world use.
Conexant, you scum!