This section shows a detailed information from Neuralink chip operation modeling, in terms of hardware and software functionality. It will detail starting from the “macro” to the micro way point of view.
The first Neuralink project is the detection. As the neuron fires, the spike (or action potentials) connects to one of the 1024 electrodes channels. The pulse needs to flow into a pass band filter (instead of a time domain, there is a frequency domain, so there is a band width frequency range of the neural spike that needs to remain) to exclude noise signals.
For a better manipulating, this filtered signal needs to be amplified digitized by an analog to digital converter (ADC). After processing it on internal microprocessor, the signal must be converted in digital to analog (DAC) and, finally, amplified be wireless transmited, as Neuralink engineers confirm, Bluetooth low energy radiofrequency (2.4 GHz).
The following image illustrate a simplified flowchart (from the author’s point of view) for the signal and the small idea of BCI initial process (neural detection and stimulation process).
Some additional blocks could be included in the chart. For example, in the process before the transmission switch, could be mixed an oscillator, with the output amplified signal, for a better signal modulation in a specific frequency. This procedure allows a better telecommunication signal processing, in terms of power and signal-to-noise factor. Or even a conditioning circuitry block for power supply. For simplicity, are not included such specifications on the flowchart.
All the process in the flowchart in terms of sensing, amplification, conversion, and processing, are totally integrated on N1 chip of the Neuralink. N1 is a totally customized ASIC (Application Specific Integrated Circuits) for the brains specifications, assuming that there is nothing on market that provides such features, as neural amplifiers f.e.
The duration of an action potential is one millisecond. N1 digitizes it at 20 kHz, so the hole signal is divided in 20 pieces to process.
The ADC divides the magnitude of each piece in 1024 levels, so N1 has 10 bits of resolution (as the digitized conversion is only about ‘1’ and ‘0’, total bits is 2¹⁰ = 1024). The detection process duration is, 900 nanoseconds (0.0000009s), and this is determined by the internal clock of the N1. Considering neural speed communication, everything on Link happens faster than the brain!
The entire Neuralink measures 23 x 8 mm. Going inside of N1 SOC (System-On-Chip), that has 5 x 4 mm, it’s possible to find 1024 channels for all electrodes. Each channel has a noise correction factor of 7.2 micro Volts. For each thread channel there is a analog pixel with 6.6 micro Watts of power consummation.
Over time, Neuralink improved three version revisions of the analog pixel, in size, power consumption while maintaining its performance. The last version on the right, (see figure below), is five times smaller than the known state of the art. Each pixel is dedicated to each electrode, as published on academic literature.
Let’s zoom in more, going from the analog pixel channel thru the thread with the electrodes. A hair has a 100 microns of diameter, the thread called linear edge, has the electrodes, whereas each one has 5 microns! Neuralink made over 20 designs of the threads, and the goal was to increase the number of the electrode at each channel, without significantly increase the width of these threads at the base (see figure below).
However, as the number of electrodes increases, these raw digital signals become too much information to upload with low power devices, and, in this case, N1 essentially saturates in signal processing, and may cause wrong measurements and overheating f.e.
The surgery robot must precisely do the incision and implant exactly on specific spots that do not harm the bloodtream.