6 min read
Updated September 13, 2026
Density is not a number to maximise. It is spatial oversampling: enough sites, close enough together, that one neuron lands on several of them at once. That redundancy is the whole point — a sorter separates two cells by how differently they appear across neighbouring sites, so the question is never "how many channels" but "are my sites closer together than the thing I am trying to tell apart."
Research use only. NeuroNexus products and the procedures described here are for non-clinical research use, and must not be used in human or veterinary medical procedures.
Everything below follows from that, including the parts that are inconvenient.
What sets density#
Two numbers, and only one of them is the one people quote.
Site pitch is the distance between neighbouring recording sites. It decides whether a neuron appears on more than one site, and therefore whether it can be isolated. This is the number that determines what you can resolve.
Channel count is how much tissue you cover at that pitch. It decides how much of the circuit you see, and it is what sets your data rate, your storage bill and your sorting time.
A design with 1,024 sites at a coarse pitch is a wide survey. A design with 256 sites at a fine pitch is a careful look at one place. Neither is "higher density" in the sense that matters until you say which of the two problems you have.
| Family | Sites | Pitch | Active length | Shanks |
|---|---|---|---|---|
| SiNAPS 1S-256 | 256 | 29.6 µm | 3,754 µm | 1 |
| SiNAPS 4S-1024 | 1,024 | 29.6 µm | 3,754 µm per shank | 4 at 560 µm |
| SiNAPS 8S-1024 | 1,024 | 30.3 µm | 1,911 µm per shank | 8 at 300 µm |
| Passive silicon | 4–256 | 20–500 µm | by design | 1–16 |
Site pitch and length are measured from the probe descriptors, not quoted from a datasheet. The passive range is the catalog's span; a specific design is the number that matters.
What density costs#
This is the part that is usually discovered rather than planned. Data rate is channel count times sample rate times the stored sample width, and it does not care how interesting the experiment is.
| Configuration | Rate | Per hour | Per 8-hour session |
|---|---|---|---|
| 64 ch at 30 kHz | 3.8 MB/s | 14 GB | 111 GB |
| SiNAPS 256 at 20 kHz | 10.2 MB/s | 37 GB | 295 GB |
| 384 ch at 30 kHz | 23.0 MB/s | 83 GB | 664 GB |
| SiNAPS 1,024 at 20 kHz | 41.0 MB/s | 147 GB | 1.18 TB |
At 16-bit stored width. Three consequences worth planning for rather than absorbing:
- A dense cohort is a storage project. Ten animals at 1,024 channels for two hours each is about 3 TB before you have analysed anything. Budget the disks with the probes.
- Sizing follows throughput, not channel count. The requirement is set by the product of channels and sample rate — see Radiens System Requirements.
- Streaming beats loading. Radiens reads a recording from disk rather than into memory, so session length is bounded by storage rather than RAM. At these rates that distinction stops being an implementation detail.
Read the quality before you trust the session#
A dense probe fails gradually. A handful of poor sites is normal; a shank or a contiguous block behaving differently from its neighbours is not, and it is the difference between a quiet region and a broken one. Because the sites are redundant by design, the pattern across neighbours tells you which you have.
Check three things before the animal is committed, in this order:
- Impedance, on the bench, against each site's own baseline rather than a nominal figure. See Impedance ↗.
- The probe heat map, live. Near-uniform is what you want; structure in the map that matches the probe's own geometry — a shank, a column, a connector's worth of channels — points at the hardware, not the tissue. See Signal Metrics ↗.
- The raw traces, briefly, on a few sites spread across the array. See Monitor ↗.
For an active-pixel probe the calibration step comes first and is not optional; the SiNAPS Manual covers it.
Sorting at scale#
Oversampling is what makes sorting work, and it is also what makes it expensive: every added site is another dimension the sorter has to consider. Two practical points.
The sort is a derived dataset, not an edit. In Radiens a sort writes a separate, linked set beside the raw recording — the recording is never modified — so a sort you distrust costs you the sort, not the session.
Sort parameters are part of the method. At 1,024+ channels the difference between two threshold choices is thousands of units either way. Save the preparation as a protocol and apply it across the cohort so the animals are comparable; that is what Curate ↗ is for.
Keeping a dense study reproducible#
Density multiplies every reproducibility problem you already had. Three habits that survive review:
- Identical preparation across the cohort. One saved protocol, applied to every animal, rather than a filter chain rebuilt from memory each time.
- Derived results stay linked to their source. Sorts, metrics and curated datasets each point back at the recording they came from, so a figure can be traced to a session.
- Record what you excluded and why. Dropped channels are a methods statement. A dense recording where the exclusions are undocumented is difficult to defend and impossible to repeat.
Choosing an acquisition path#
The probe and the software are separable decisions, and it is worth making them separately.
| If your lab… | The straightforward path |
|---|---|
| Runs SiNAPS and nothing else | SmartBox Pro with Radiens, or the vendor's own acquisition system — both record the probe properly |
| Runs SiNAPS alongside passive arrays, grids or Neuropixels | Radiens, so every recording lands in one workspace on one probe-geometry model |
| Needs stimulation as well as recording | XDAQ with an SR-series headstage — active-pixel probes are record-only |
| Already owns another acquisition system | Radiens reads it; the software choice is not locked by the hardware you bought |
The last row is the one that saves money. A lab that has already bought an acquisition system does not need to replace it to change what it can get out of the recordings.