929 Neuromorphic Processor (SNN)
929 : Neuromorphic Processor (SNN)
- Author: Duarte Monteiro
- Description: Time-Multiplexed SNN accelerator
- GitHub repository
- Open in 3D viewer
- Clock: 50000000 Hz
How it works
NeuroPulse is a 4-neuron spiking neural network accelerator using the Leaky Integrate-and-Fire (LIF) model. Weights are held off-chip by an RP2040, streamed in every pass via PIO — keeping the transistor budget under 2000T.
Architecture:
- The FSM processes one neuron at a time, 7 states each → 28 cycles per full network pass
- At 50 MHz this gives ~1.78M passes/second — far beyond the ~100 Hz needed for biological-speed simulation
- Per-neuron state: 6-bit membrane potential (
vmem) and 3-bit activity trace - Leak:
vmem = vmem - (vmem >> 2)— 75% retention per cycle, no multiplier needed
FSM sequence (per neuron):
| State | Action |
|---|---|
| LATCH | Capture ui_in[7:4] as external spike injection (neuron 0 only) |
| TRACE | Decay or reset this neuron's activity trace |
| ACCUM_0–3 | Read weight[N][i] from RP2040 on ui_in[3:0], accumulate spike[i] * weight |
| THRESH | Apply threshold, emit spike, reset vmem if fired |
Hebbian learning (LTP only): when a neuron fires with learn_ena=1, ltp_pulse goes high on uo_out[7].
The RP2040 watches this signal and increments the corresponding weight in its own table.
How to test
Connect an RP2040 running the PIO weight-streaming firmware (see docs/rp2040.md).
Minimal smoke test without RP2040:
- Hold
rst_nlow, then release. Allvmemandspikesclear to zero. - Set
ui_in[7:4]to inject external spikes (e.g.0001= fire neuron 0). uo_out[3:0]shows spike outputs; watchuo_out[4:5]to see the FSM stepping through neurons 0–3.uio_out[6:5]shows the active synapse during ACCUM states;uo_out[6](in_accum) gates when the RP2040 must present weights.
With RP2040:
- Pre-load
WEIGHTS[4][4]table on the RP2040. - RP2040 PIO streams
weight[N][syn]onui_in[3:0], timed to the chip's deterministic FSM. - Monitor
uo_out[7](ltp_pulse) +uo_out[5:4](neuron_sel) for Hebbian weight updates.
External hardware
- RP2040 (e.g. Raspberry Pi Pico): holds the 4×4 weight table, streams weights via PIO, applies LTP updates.
See
docs/rp2040.mdfor the full interface protocol and PIO program.
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | Weight bit 0 (from RP2040, valid during ACCUM) | Spike output neuron 0 | learn_ena (input: 1 = Hebbian learning on) |
| 1 | Weight bit 1 (from RP2040, valid during ACCUM) | Spike output neuron 1 | in_accum (high when RP2040 must present weight) |
| 2 | Weight bit 2 (from RP2040, valid during ACCUM) | Spike output neuron 2 | Neuron select bit 0 — neuron_idx[0] |
| 3 | Weight bit 3 (from RP2040, valid during ACCUM) | Spike output neuron 3 | Neuron select bit 1 — neuron_idx[1] |
| 4 | Ext spike neurons 4 (latched at S_LATCH, neuron 0 turn) | Spike output neuron 4 | Neuron select bit 2 — neuron_idx[2] |
| 5 | Ext spike neuron 5 | Spike output neuron 5 | Synapse select bit 0 — syn_idx[0] (valid when in_accum=1) |
| 6 | Ext spike neuron 6 | Spike output neuron 6 | Synapse select bit 1 — syn_idx[1] (valid when in_accum=1) |
| 7 | Ext spike neuron 7 | Spike output neuron 7 | Synapse select bit 2 — syn_idx[2] (valid when in_accum=1) |