Skip to content

Neuromorphic hardware

The BrainScaleS-2 tab of the Stimulation setup pane (View → Show Stimulation setup) delivers the current input array to the BrainScaleS-2 (BSS-2) accelerated neuromorphic chip and adds the response as a new data group. The chip cannot inject arbitrary currents, so input is either a constant current per segment or spikes from a virtual source population.

Verified on hxcube7fpga0chip57_1 via quiggeldy.bioai.eu:11750 (October 2026): in a 10-neuron staircase each neuron fired only during its own pulse.

How it works

The BSS-2 software (pynn_brainscales) exists only in the EBRAINS container. Connect starts resources/bss2/neuroseq_bss2_helper.py inside it with singularity exec, and NeuroSeq talks to it over JSON lines. The helper reaches the chip through quiggeldy (the BSS-2 job server). One connection is shared by all project tabs; each tab receives only its own results.

Setup

  1. Container image. Lab machines: /containers/stable/latest (default) with Container app dls. Elsewhere pull the EBRAINS image and pick the .sif with Browse… (leave the app empty):
mkdir -p tmp
SINGULARITY_TMPDIR=$PWD/tmp singularity pull \
  docker://docker-registry.ebrains.eu/esd/master/image:lab_latest.oci
  1. Setup. quiggeldy.bioai.eu ports 11717 (stable), 11750 (experimental, default) and 11780 (special). Only allow-listed IPs can connect; ask the BSS-2 team to add yours. With lab_latest only experimental worked (October 2026); the others fail with Could not submit request. Refresh also downloads the EBRAINS setup list. Host, port and expected Identifier stay editable.
  2. Token (optional). Not needed from an allow-listed IP. Otherwise paste clb_oauth.get_token() output, or set QUIGGELDY_TOKEN. The token stays in memory only, is passed through the environment, and is masked in the log.
  3. Connect. Connected chip shows the identifier (with a warning if it differs from the expected one).

Settings except the token are remembered.

Start-up time

  • Without squashfuse, each start unpacks the .sif (about 1.5 min). Build a sandbox once and select that folder instead (starts in about 1 s):
singularity build --sandbox image_lab_latest_sandbox image_lab_latest.oci.sif
  • The first activation of the container's spack environment takes about 50 s; it is then cached per image (~/.cache/neuroseq/bss2 on Linux).

Input modes

Neurons (auto: one per input row) and Max neurons (512, one chip). Row i drives hardware neuron i; extra rows are dropped and missing ones get no input, with a warning. A single-row array drives every neuron.

Constant current (input array) (default). The array is split into segments where any row changes value. Each segment is a separate hardware run starting from rest (chained runs did not work on the chip), with

constant_current_i_offset = round(clip(value / Stim value at max, 0, 1) × Max i_offset)

and the current enabled where the value is above 0. Max i_offset: 500 (range 0-1022). Stim value at max: array maximum. Each segment costs about 0.5 s of hardware time; more than 1000 segments trigger a warning.

Spike input. A SpikeSourceArray (one source per neuron) with one-to-one excitatory synapses of Weight 63 (0-63). Spike source:

  • From input array (threshold): sample the array every Event interval (20 ms biological) and spike where it exceeds Threshold (0).
  • From project spike data: a spike dataset from a data group; neuron i drives hardware neuron i.

Time base

BSS-2 runs about 1000x faster than biology. Speed-up (default 1000):

  • hardware ms = biological s × 1000 / speed-up (1 s of input is a 1 ms run),
  • biological s = hardware ms / 1000 × speed-up.

Delivery and readback

Deliver current input array shows a summary and asks for confirmation. Neurons are HXNeurons with the chip's nightly calibration. The result is a data group BSS-2 <stim name> <date time> (undoable) with:

  • BSS-2 spikes of all driven neurons, and
  • BSS-2 membrane voltage (neuron n) for the neuron in Record v of neuron (one MADC, so one neuron per run), resampled to a regular step, raw units.

Stop kills the helper; a run already on the chip is discarded and you must reconnect. A running delivery keeps the project tab from closing.

Rheobase routine

In Design dynamic input → rheobase, choose BrainScaleS-2 (neuromorphic). Connect in the dialog uses this tab's settings.

  • Each probe is delivered as constant current with the tab's settings. Probe amplitudes are fractions (0-1) of Max i_offset.
  • The probed neuron's membrane voltage is recorded. All probes go into one data group, BSS-2 rheobase <date time>, aligned with the input.
  • Every probe starts from rest; spikes after the current ends are not recorded.
  • The chip is noisy near threshold: Repeats per probe (3 on hardware targets) counts a response when most runs spike.
  • If the maximum cannot make a short pulse fire, that duration is marked (red triangle); the chronaxie search continues only if the maximum is at least twice the rheobase. Raise Max i_offset (up to 1022) for more current.
  • Stop ends the search without killing the shared helper.

Debug log

The Debug log lists every step (launch command, environment, helper versions, and with Show protocol traffic every message). The token never appears. Copy log copies it with version, host and settings for bug reports; Clear empties it.

Chip notes (October 2026)

  • pynn.run() durations are hardware ms; runs need pynn.reset() in between.
  • Constant current requires HXNeuron (not CalibHXNeuronCoba) on the nightly calibration.
  • Chained APPEND runs executed only the first part, hence one run per segment.
  • The membrane is sampled about every 32 ns hardware time (32 µs biological at 1000x).
  • Neurons differ: at i_offset 500 they fired 2-9 spikes per 100 ms pulse.