Stochastic & Neuromorphic Computing  /  Explore  /  Neural codecs

Neural signal codecs
fitting a brain into a radio

A modern brain–computer interface reads a thousand electrodes at tens of kilohertz — hundreds of megabits per second — through a wireless link that carries a small fraction of that. The gap is the binding constraint of implantable neurotech. SC-NeuroCore closes it with lossless neural codecs that exploit how sparse and structured spike data is, and a mode that is bit-true to on-implant hardware.

The bandwidth gap

The numbers are stark. Raw electrode data out-runs the uplink by one to two orders of magnitude, and the radio is fixed by power and thermal limits inside the skull — so the data has to shrink, losslessly, before it leaves.

SystemChannelsRaw rateUplinkGapCodec
Neuralink N11,024205 Mbps10–20 Mbps10–20×WaveformCodec (~25×)
Neuralink next-gen3,072614 Mbps10–20 Mbps30–60×WaveformCodec
Neuropixels 2.038477 MbpsarchivalstorageWaveformCodec
Closed-loop BCI256–102451–205 Mbpson-chip< 1 msStreamingCodec
Live — will it fit the uplink?
compressed Mbit/sraw Mbit/s · ratio

Raw rate assumes 30 kHz × 10-bit sampling per channel; the codec ratio follows the measured ISI curve (hundreds-fold on very sparse trains, single-digit on dense ones) or the ~24× hybrid WaveformCodec. Illustrative of the trade — exact ratios depend on the recording.

01  WaveformCodec — compressing the raw electrode

The headline codec is a hybrid that splits the raw signal into what matters where: the spike band is reduced to spike times and coded losslessly (inter-spike intervals + Huffman), while the slow background field potential is down-sampled, delta-coded, quantised and entropy-packed. Measured around 24–25× on realistic recordings — enough to drop a 1,024-channel Neuralink-class stream into a Bluetooth uplink.

raw electrode ── spike band → ISI + Huffman (lossless) └─ background LFP → 4× downsample + delta + quantise + zlib
Deeper: lossless where it counts
Spike timing is where the information a BCI decoder needs actually lives, so that band is kept bit-exact; the background field is perceptually and analytically less sensitive, so it tolerates the lossy stages that win most of the ratio. On a 1,024-channel stream the compressed output lands around 9% of a 15 Mbit/s Bluetooth budget — the difference between an implant that can transmit and one that cannot.
02  Spike codecs — exploiting sparsity

When the data is already a spike raster, six codecs each attack its structure differently. Because spikes are rare, the achievable ratio scales inversely with firing density — enormous on quiet trains, single-digit when everything is firing at once.

$$\text{ratio} \approx \frac{\text{raster bits}}{\text{event bits}} \;\downarrow\; \text{as density} \uparrow, \qquad \text{ISI: } \sim\!400\times \text{ at } 0.1\%,\ \sim\!9\times \text{ at } 30\%$$
Deeper: the codec family, and picking one
The ISI codec (inter-spike intervals + auto-selected varint/Huffman entropy) is the general-purpose baseline; the predictive codec adds a Markov context model with XOR residuals and wins roughly 3× over ISI on bursty, structured data (~25× on bursty trains); delta, AER (address-events) and streaming codecs trade ratio for latency or for the bus format a neuromorphic core expects. A recommend_codec(n_channels, firing_rate, latency_ms) helper picks one from the constraints, and all of them round-trip losslessly — the spike-codec notebook on this site benchmarks them across densities.
03  On-implant — bit-true to silicon

Compression only helps if it can run inside the implant, on a power budget of milliwatts. The predictive codec has an LFSR mode in Q8.8 fixed-point that is bit-true to generated Verilog — the same compression, provable on the same silicon path as the rest of the toolkit, so the codec is ASIC-deployable rather than a host-side afterthought.

04  Closed-loop & latency

A therapeutic closed loop — read, decode, stimulate — cannot wait for a big buffer to fill. The StreamingCodec compresses incrementally with a bounded per-window latency (sub-millisecond targets), trading a little ratio for the timing a control loop needs, while the archival path uses the higher-ratio block codecs.

Reference — which codec
CodecBest forMethodMeasured ratio
ISIgeneral purposeinter-spike intervals + varint/Huffman~400× at 0.1%, ~9× at 30%
Predictive (context)structured / burstyMarkov context + XOR residual~25× on bursty (≈3× over ISI)
Predictive (LFSR)on-implant hardwareQ8.8 LFSR, bit-true Verilogsame ratio, ASIC-deployable
WaveformCodecraw electrode streamshybrid spike + LFP~24–25×
StreamingCodecclosed-loop < 1 msincremental, bounded latencylatency-bounded

Compression figures are measured on specific recordings and densities and are directional, not guarantees; every codec here round-trips losslessly (except the WaveformCodec's background-LFP stage, which is deliberately lossy). System bandwidths cite published device specifications.