Neuronova SDK — Early Access

This is an early-access release of the Neuronova SDK, designed to give you hands-on access to a new computing paradigm: always-on, ultra-low-power analog AI at the edge.

Some features may be incomplete, APIs may evolve, and parts of the experience will change as the technology matures. This is expected at this stage.

We’re releasing the SDK early on purpose. We believe the fastest way to build the right platform is to build it together with the people who are pushing real application to the real world, not in isolation.

Your feedback is extremely important to us. What works, what doesn’t, what feels missing, all of it directly influences our roadmap, our tooling, and the next iterations of both software and silicon.

This is not just about testing an SDK.
It’s about shaping a new always-on intelligence layer at the edge.

Thank for being with us!

OK

Hardware-aware Losses

These losses encode constraints and non-idealities of the Neuronova neuromorphic chip. They are designed to regularize network parameters so that trained models can be deployed on hardware.

These losses are optional and composable. They represent one practical toolkit developed at Neuronova, not a universal prescription.

---

Hardware Constraints Recap

The chip imposes several constraints, some can be approached during training with penalty losses that alter complexity landscape in order to let the model reach a mappable configuration thorugh optimization. Those type of constraints are:

• Bounded weight range:
  All weights must lie in the interval \([-0.9, 0.9]\)

• Structured sign topology:
  Groups of neurons must share aligned weight signs, reflecting physical routing constraints

---

Weight Magnitude Loss

Function: weight_magnitude_loss

Penalizes weights whose absolute value exceeds a specified limit.

Import with:

from nwavesdk.loss import weight_magnitude_loss

Description

For each module in the model:

• All standard weights and recurrent weights are inspected
• Values outside the allowed range are softly penalized
• The penalty is normalized by the total number of parameters

This loss is size-invariant and does not bias larger models.

Mathematical Form

For each weight \(w\):

\[ L_w = \max(|w| - \text{limit}, 0)^2 \]

The final loss is the average over all parameters.

Arguments

• model: PyTorch module
• limit: Maximum allowed absolute weight (default: 0.9)

Returns

A scalar tensor representing the normalized penalty.

---

Topology Loss

Function: topology_loss

Encourages weight sign alignment according to hardware topology constraints.

from nwavesdk.loss import topology_loss

Description

Weights are grouped in non-overlapping blocks of 5 neurons. Within each group, incoming weights are encouraged to have consistent signs.

This reflects:

• Routing constraints of the chip
• Limitations on mixed-sign fan-in

The loss penalizes sign disagreement within each group.

Behavior

• Perfectly aligned signs → zero penalty
• Mixed signs → increasing penalty
• Works on both feedforward and recurrent weights

Arguments

• model: PyTorch module
• lam: Scaling factor for the topology regularizer

Returns

A scalar tensor proportional to the topology violation penalty.