1000 samples/sec, 4 bytes per sample, 32 nodes - that's 128Kb / second. Not crazy by any means. However, you then have to process that... That's doing an FFT + binning (in the case of brain waves). All doable at probably 60 FPS (where a frame is the prior 1 - 3 seconds).
Still not a problem, then try to do error correction! You have 32 sensors, what's real, what's an artifact? That's when things slow way down typically. This type of analysis can be done after the fact and you essentially guess at errors. Removing known issues or more likely from a research perspective they just mark those sections of the data and remove them.
Here's where what we built comes in:
http://synaptitude.me/blog/using-computer-vision-to-improve-...
The above is all about post processing, however, I managed to improve the method and write a proof-of-concept for a real-time application that removes artifacts in EEG signals, BUT leaves the underlying signal (AKA you don't need to throw away the whole sample). This can function to capture error corrected brain waves at 50 Hz with 512 samples per second + a 50hz video stream. Probably could do more.
This is extremely technically complicated and probably took a year or so to develop. This enables real-time stream (50Kb / second) to our hosted server and real-time applications:
http://synaptitude.me/blog/view-brainwave-data-thinksuite-io...
Unfortunately, I couldn't find someone willing to support the work further, and even though we had maybe 30 people vouching $300 each. That wasn't enough to continue the work. Most of this work concluded in early 2017.
For reference, I may apply for an SBIR grant coming up to fund additional work on this (my wife started graduate school, utilizing EEGs, so it aligns pretty well).
Honestly, if anyone is interested in this please let me know. Its an area of interest, just couldn't justify the expense or time investment (now working on this: https://hnprofile.com/)