it must require:
very high speed sampling rates.
So high we'll use >100x data rates and storage.
this is to have ultra-low latency for tracking and live work.
most likely this would not run on a standard computer-based system very well due to driver latency. standalone embedded systems would provide best performance.
the systems require end-to-end very high sample rates, from ADC to DAC, in order to achieve very low latencies far below the ability to hear it. simple integral upsampling provides backward compatibility of media assets.
current practice zero-latency analog monitoring will remain essential, allowing favored outboard analog inserts, but ultrasampling will add near-analog latency, allowing favored processing in digital.
sampling rates required are one to two orders of magnitude greater than present to even begin to approach near-analog latencies, more specifically, below perceptual threshold. many claim 7-10ms being below threshold but most musicians can hear as low as 1ms latency, especially percussionists, when monitoring via headphones.
this all works because uncompressed video stream servers, data rates and storage can handle the bandwidth such ultrasampling techniques. the infrastructure is already here.
ironically, this will take some time as it obviates all existing gear. however, we've gone through that before and would be arrogant to say it won't happen again.
Examplea typical ADC takes 7-15 samples to convert into digital (delta-sigma type). the DAC takes 3-4 samples to convert into analog.
so let's assume the worst: 19 samples, plus 1 more for a buffer internally. let's double it for devil's advocate, for a total of 40 samples. at 44.1kHz, that's nearly 1ms. at 10x ultraSampling, 441kHz, latency is 0.1ms.
Assuming each patch in a signal flow takes 1 sample on either side, plus 1 sample for processing (no lookahead devices in this scenario), each process in a signal path takes 3 samples. Assume we have 8 processes; this is a worst-case conservatism. Such discrete processes might be Fader, 6 inserts, and Output to bus, for a total of 24 samples. Add the devil's advocate 40 samples for AD/DA path, for a total of 64 samples. Ultrasampling at 10x (441kHz) would reduce complete analog-in to analog-out from 1.5ms to 0.15ms.
To approach analog latency, ultraSample at 100x (4.41MHz) and such latency would be 0.015ms, or less than 1 sample at existing 1x rates (44.1kHz single sample is 0.023ms).