I confirm what @Johan says here
Although the DSP is optimized for real-time processing - which means - without introducing latency, no DSP in this world has the computing power needed to perform the complex calculations of separation algorithms in real time. They are not designed for that at all.
In the best case they can “process” an audio signal, apply a filter (hi pass/low pass) and some effects (reverb, delay, etc…). This kind of “processing” is a joke for a DSP compared to the computing power needed to perform a separation in real time. It is not comparable to the tasks for which DSPs are designed.
Consider that an M1 processor filled with NPU cache already takes a while to do this separation on a desktop, so a DSP simply can’t do it, let alone in “real time”.
DSPs are designed to offload the CPU to do simple tasks with little latency.