SSOS ESP32-S3: reproducible multi-board language inference and split-training packages #12841
tylorsaling-source
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hello Espressif and Arduino-ESP32 community,
I would like to share SSOS ESP32-S3, an independent open-source experiment built on ESP32-S3-WROOM-1U N16R8 boards with the Espressif Arduino-ESP32 core:
https://github.com/tylorsaling-source/SSOS-ESP32
The physically tested V3 package splits a pinned 465,504-parameter causal-language model across ESP32-S3 boards connected over 40 MHz SPI. The accepted two-board run used two independent interleaved contexts and produced 240/240 expected tokens at a median 44.730763 aggregate tok/s. The explicitly named Paired Two-Board Non-Interleaved Baseline was 18.345216 tok/s, so the measured interleaving gain was 2.438279x. This is an aggregate-throughput scheduling result, not a claim of 2x lower single-token causal latency or a general 2x model speedup.
The optional one-master/two-worker setup uses two independent SPI/DMA lanes and four model contexts. Its accepted gate produced 96/96 expected tokens at 87.927387 aggregate tok/s.
V4.0.0 and V4.0.1 are separate downloadable 549,984-parameter split-training lineages. Each package includes its own checkpoint and integrity boundary; the detached master/worker reference gate matches continuous training with zero loss, gradient, and updated-weight difference.
The repository includes exact role-specific source and binaries, Arduino-ESP32 versions, wiring, model artifacts, checksums, raw machine-readable results, recovery checks, and fail-closed oracle verification. I would especially welcome review or reproduction feedback concerning:
Relevant releases:
This is an independent community project and is not affiliated with or endorsed by Espressif Systems or Arduino.
All reactions