454 Spiking Neural Network WTA Inference Engine (GF180)
454 : Spiking Neural Network WTA Inference Engine (GF180)
- Author: Prof. Santhosh Sivasubramani, IIT Delhi
- Description: GF180mcuD digital spiking neural network inference engine: 4 LIF neurons with programmable 4x4 signed synaptic weight matrix, per-neuron bias, programmable global threshold/leak/refractory, hard winner-take-all lateral inhibition, optional self-recurrence, SPI-configurable 16-bit register file, 4-channel 2-bit input current coding, 4 spike outputs, 2-bit winner ID output, any-spike and busy flags; deterministic single-cycle neuron update
- GitHub repository
- Open in 3D viewer
- Clock: 25000000 Hz
How it works
A 4-neuron digital Spiking Neural Network (SNN) inference engine with hard Winner-Take-All lateral inhibition for TinyTapeout GF 26a (1×2 tile).
Each clock tick (rate set by an 8-bit tick divider) every non-refractory neuron
accumulates a MAC of its bias, the dot product of the 4-channel 2-bit input
current vector with its row of the 4×4 signed weight matrix, and an optional
self-recurrence term. The resulting increment is added to the 16-bit signed
membrane (saturating on overflow). If the updated membrane crosses the global
threshold the neuron fires a 1-tick spike and resets; if hard_wta=1, the
lowest-index firing neuron is the sole winner and all other firing neurons are
forced into the refractory_cnt inhibition window for refractory cycles.
All datapath parameters (threshold, leak, refractory, tick_div, bias, weights, recur_gain) are programmable over the SPI register file. Membrane and refractory counters are memory-mapped for debug.
How to test
- Reset (
rst_n=0≥ 10 clocks). - Program via SPI:
- threshold, leak, refractory, tick_div, recur_gain
- per-neuron bias[0..3]
- weight matrix w[0..15]
- Write
ctrl = {recur_en, hard_wta, 0, global_en=1, irq_en}to start. - Drive
ui_in[7:0]with the 4×2-bit input current vector. - Observe spikes on
uo_out[3:0]and winner ID onuo_out[5:4];uio_out[5]pulses every tick for oscilloscope triggering. - Poll
status (0x20)for latest spike vector + sticky overflow flag.
External hardware
- An SPI master (any Raspberry Pi / microcontroller) for register programming.
- A 4×2-bit input pattern generator on
ui_in[7:0]— could be raw GPIO, a small logic-analyser pattern, or an external event source. - Oscilloscope or logic analyser on
uo_out/uio_outfor spike capture.
Register summary
| Addr | Name |
|---|---|
| 0x00 | ctrl |
| 0x01 | threshold |
| 0x02 | leak |
| 0x03 | refractory |
| 0x04 | recur_gain |
| 0x05 | tick_div |
| 0x08..0x0B | bias[0..3] |
| 0x10..0x1F | w[0..15] |
| 0x20 | status |
| 0x30..0x37 | membrane lo/hi (read-only) |
| 0x38..0x3B | refractory_cnt (read-only) |
ctrl bits
| Bit | Name | Meaning |
|---|---|---|
| 0 | irq_en | gate uio[7] IRQ pulse on any spike |
| 1 | global_en | master enable (also drives busy output) |
| 2 | learn_frozen | reserved for future on-line STDP |
| 3 | hard_wta | 1 = lateral inhibition of losers; 0 = soft WTA |
| 4 | recur_en | 1 = add recur_gain when fired_prev[i]=1 |
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | in_cur0[0] | spike[0] | spi_cs_n |
| 1 | in_cur0[1] | spike[1] | spi_mosi |
| 2 | in_cur1[0] | spike[2] | spi_miso |
| 3 | in_cur1[1] | spike[3] | spi_sck |
| 4 | in_cur2[0] | winner_id[0] | overflow_ever |
| 5 | in_cur2[1] | winner_id[1] | tick |
| 6 | in_cur3[0] | busy | winner_valid |
| 7 | in_cur3[1] | any_spike | irq |