Docs / Quickstart

Quickstart

Install NMC, compile a PyTorch model to .snn, and run inference on the NT3000 eval kit. Estimated time: 10–15 minutes.

1. Installation

Prerequisites: Python 3.9–3.12, pip, and an NT3000 eval kit connected via USB-C. The eval kit drivers are bundled with the NMC CLI.

install

# Install NMC CLI
$ pip install nmc-cli
Successfully installed nmc-cli 0.9.1
$ nmc doctor
✓ Python 3.11.3
✓ NT3000 eval kit detected (FW 1.2.0)
✓ USB-C driver OK
All checks passed

If nmc doctor reports no device, check the USB-C connection and ensure libusb is installed (brew install libusb on macOS or sudo apt-get install libusb-1.0-0 on Linux).

2. Export your model

NMC accepts TorchScript (.pt) or ONNX (.onnx) exports. For the quickstart, use the bundled sample model:

export.py

import torch
from neurmorph.samples import GestureNet

model = GestureNet(num_classes=12)
model.eval()
example = torch.zeros(1, 9, 100)
traced = torch.jit.trace(model, example)
traced.save("gesture_net.pt")

3. Compile to .snn

Run the compiler with the --calibrate flag to run threshold calibration before serialization. The calibration pass uses the 512-sample default dataset bundled with NMC.

Basic compilation

compile

$ nmc compile gesture_net.pt \
  --target nt3000 --calibrate
→ pass 1/6 dead-spike elimination
→ pass 2/6 temporal op fusion
→ pass 3/6 rate normalization
→ pass 4/6 threshold calibration
→ pass 5/6 HAL mapping (nt3000)
→ pass 6/6 binary serialization
output: gesture_net.snn (87 KB)
accuracy delta: -1.4% vs baseline

Inspect the binary

Use nmc inspect to view the binary metadata, memory layout, and expected SRAM usage before deploying.

inspect

$ nmc inspect gesture_net.snn
target:  nt3000
neurons: 4096
sram:    87 KB binary + 9 KB state
layers:  3
nmc:     0.9.1

4. Deploy to eval kit

Flash the compiled .snn binary to the NT3000 eval kit over USB-C. The eval kit firmware handles loading and starting the runtime loop.

deploy

$ nmc deploy gesture_net.snn \
  --device nt3000:usb0
→ connecting to NT3000 (FW 1.2.0)
→ flashing 87 KB
→ verifying checksum
Model deployed. Runtime active.

5. Verify inference

Send a test sample via USB-C and read the classification result. The eval kit firmware exposes a simple JSON-over-serial protocol.

verify

$ nmc infer --device nt3000:usb0 \
  --sample samples/swipe_right.npy
class:    swipe_right (idx 7)
latency:  0.11 ms
power:    38 mW (active)
Inference OK

Next steps

Read the API Reference for full NMC SDK and CLI documentation.
See Supported Hardware to port to BrainPulse-2 or SynCore-V.
Use Compiler docs to understand the six optimization passes.
Contact the engineering team for pilot program access and direct support.