Pi 3B · 1GB RAM layer farm

What I was poking at

Experiments 01–03 asked how cameral width, train modes, and volatile cloud behave under Lucy. Experiment 04 takes the stack somewhere meaner: Raspberry Pi 3B-class boxes with about a gigabyte of RAM, and asks whether Welvet’s layer catalogue can actually train there — not as a slogan, as finished Tide boards.

Three hosts split the work. pi-240 took the spatial / dense family (cnn1–3, convt1–3, dense). pi-241 took sequence and attention-ish layers (embedding, gdn, lstm, mamba, mha, rnn, swiglu). pi-243 took norms and composites (kmeans, layernorm, metacognition, parallel, residual, rmsnorm, sequential, softmax). Overview pack: training-modes research, a matched mode-vs-SGD report across all 22 layer boards, and a 3D-grids audio briefing.

The setup, in plain terms

Each layer got its own pi_collection Tide sprint: dtype × train mode × arch on SIMD, recorded as a Lucy PDF plus checkpoint history/progress JSON. Most boards finished 2346 / 2346 cells in epoch 1. Two were still chewing at extract — mamba ~86% (2024/2346) and softmax ~94% (2197/2346) — which is honest edge reality, not a hidden failure of the others.

Lucy Score stays Throughput × Availability × Acc / 10,000. The mode-vs-SGD report is the cross-cut that matters: every non-SGD mode compared to SGD on matched cells (same dtype/format/arch/backend). Across 22 layers and ~48.9k global board cells it logged hundreds of thousands of win events. Headline rates: float32 Acc beat-rate 98%, quant/low-bit Acc beat-rate 97%, quant/low-bit Score beat-rate 81%. SGD is the reference row, not the expected champion.

What showed up

The farm ran. That alone answers the “can you train Welvet on a 1GB Pi?” question with artefacts, not vibes. Dense on pi-240 was the live-fit rocket: Lucy Score ~16344 on fp4 tween_chain, with throughput north of 100k outputs/sec on the champion cell. convt1 and the metacognition / residual / embedding boards also landed in the thousands. CNN1 even posted ~94% hard Acc on its Lucy champ (float32 tween_chain) — a reminder that Acc champions and Score champions are related but not identical stories.

Heavy layers did not refuse to boot. MHA, LSTM, Mamba, and GDN produced finished (or nearly finished) boards with tiny Lucy Scores in the ~21–29 range — slow, tight Availability, still ticking cells on the same class of hardware. That is the edge plot twist: the exotic layers are expensive, not imaginary. Binary and other tiny dtypes keep showing up on hard-Acc / mobile-trap axes; Tween, TweenChain, Split/Sparse, and occasional plain SGD trade the Lucy crown depending on the layer.

The training-modes research and mode-vs-SGD cut agree with Experiments 02–03: target-propagation / Split families are how you keep learning when RAM and duty cycle are the scarce resources. Classical BP/SGD remains useful as a baseline and sometimes wins a board (softmax, lstm, mha, convt2, kmeans, metacognition champs wore SGD) — but as a global matched comparison it loses Acc most of the time, including in low-bit storage.

What I take from it

Welvet’s layer surface is not desktop-only theatre. On ~1GB Pis you can finish dense/CNN/norm sprints, keep sequence and attention layers alive, and show statistically that non-SGD modes beat matched SGD on accuracy across the collection. Use the overview PDFs for the argument; open each pi-### · layer bundle for the raw Lucy board and checkpoints.

If Experiments 01–03 were about live-fit knobs and longevity, Experiment 04 is the proof that those knobs still turn when the machine is a Pi.