March 1, 2026
Tone wars: neural or just nerdy?
Neural Guitar Pedal – Optimizing NAM for Daisy Seed Arm Cortex-M7
Tiny pedal, big brain: awe, eye-rolls, and assembly nightmares
TLDR: They crammed a smart amp model into a tiny Daisy Seed pedal by hand-tuning math and a compact file format, hitting faster-than-real-time playback. Commenters split between cheering the low-level wizardry and asking whether it's truly 'neural' or just clever signal tricks, with meme-y assembly groans.
Guitar nerds and code wizards just watched a tiny pedal get a brain transplant, and the comments went wild. The team squeezed Neural Amp Modeler (a smart amp simulator) onto the small Daisy Seed board by hand-tuning math, swapping a heavy function for a lighter one, and inventing a lean model file called “.namb.” Translation: they made it run faster than the music itself with room for extra effects.
Dev pride splashed across the thread with one engineer flexing about “hand-rolled” math loops, while another joked about “the joys of staring at assembly,” a programmer’s version of a horror movie. One grateful onlooker chimed in with a warm “pretty neat stuff,” but the biggest eyebrow-raiser was the blunt: “What makes it Neural?” Cue the classic internet split—some celebrate the wizardry, others question whether “neural” is a buzzword or a bona fide brain.
Beyond the drama, readers clocked the real wins: a compact loader for embedded hardware, a new file format that cuts bloat without killing tone, and a path toward “Slimmable NAM,” which promises models that adapt to whatever box you throw them in. Code is coming, too. In short: tiny board, big flex, and a comment section riffing between awe, skepticism, and memes.
Key Points
- •A NAM loader was built for the Electrosmith Daisy Seed (ARM Cortex‑M7) to study embedded deployment.
- •Initial performance using an A1‑Nano model (with ReLU) was over 5 seconds of compute for 2 seconds of audio, highlighting major constraints.
- •Profiling revealed Eigen’s small‑matrix GEMM as a bottleneck; specialized routines and targeted optimizations were added.
- •A new compact binary model format (.namb) was created to replace .nam JSON on embedded devices, converted via a companion app.
- •Post‑optimization, the model processes 2 seconds of audio in ~1.5 seconds, informing A2 design and leading to “Slimmable NAM.”