aliemo/transfomers-silicon-research
Research and Materials on Hardware implementation of Transformer Model
Quick Summary
Transformer Models Silicon Research is a community‑maintained catalog of papers that map Transformer‑style neural networks (e.g., BERT, LLMs, Vision‑Transformers) onto real silicon – ASICs, FPGAs, processing‑in‑memory chips, and other custom accelerators. The repo stores the data in a single YAML file, auto‑generates a searchable static website, and runs a weekly GitHub‑Actions pipeline that pulls new entries from OpenAlex/arXiv and opens a pull‑request for human review.
What the project actually does
| Feature | How it works (as described in the README) |
|---|---|
| Paper catalog | data/papers.yaml holds ~580 entries (2024‑06 snapshot). Each entry records DOI, venue, year, hardware platform, model tags, and a three‑stage review status (Pass 1‑3). |
| Generated outputs | A Python script (scripts/generator.py) reads the YAML and produces a human‑readable README.md, a CSV export, and a simple plot. |
| Static website | website/build.py turns the catalog into a filterable HTML site hosted on GitHub Pages. Users can filter by year, venue, hardware platform, model, and review status. |
| Weekly ingest pipeline | A GitHub Action runs every Monday. It queries OpenAlex and arXiv for recent papers, filters those mentioning transformer hardware, appends them to the YAML with ignore: check and review_pass: 1, then regenerates the site and opens a PR (auto/papers-ingest). |
| Local admin UI | Run scripts/admin_server.py (after pip install -r requirements.txt) to get a tiny web UI (http://127.0.0.1:8787/admin/) where curators can accept/reject entries and advance them through Pass 1‑3. |
| Contribution workflow | Contributors add or edit entries in papers.yaml, push a PR, and the CI will rebuild the site automatically. |
| Optional AI triage | The pipeline can use Groq or Gemini APIs for a heuristic “AI‑assisted” relevance check, but it works without them. |
Who might find this useful?
- Hardware researchers looking for prior silicon implementations of transformer workloads.
- Silicon designers (ASIC/FPGA teams) who need a quick literature survey of what has already been built.
- Students or survey writers needing a curated bibliography of transformer‑on‑silicon papers.
- Tool builders who want a ready‑made dataset (
papers.yaml/ CSV) for training models that classify hardware papers.
How to get started locally
# Clone the repo
git clone https://github.com/aliemo/transfomers-silicon-research.git
cd transfomers-silicon-research
# Install the tiny Python environment
pip install -r requirements.txt
# Re‑build the website (outputs to website/dist)
python website/build.py
# Open the generated index.html in a browser
# Run the admin UI for curation
python scripts/admin_server.py # then visit http://127.0.0.1:8787/admin/
Limitations / Things to note
- The catalog is curated, not exhaustive; it relies on human reviewers to mark entries as “silicon” or ignore them.
- The auto‑ingest only looks back ~21 days by default; deeper historical pulls must be run manually (
scripts/pipeline/fetch_papers.py). - The site is static – no server‑side search, only client‑side filtering.
- No actual hardware designs or code are provided; the repo is a bibliography/metadata service.
TL;DR
A well‑organized, automatically updated bibliography of papers that implement Transformer‑family models on real hardware, with a searchable web front‑end and a simple UI for community curation.
Related
- Project
- Project
- Project
- Project