Initial report-backed release of the signed INT8 tensor accelerator.
Evidence:
- 63 / 63 exact PyTorch-to-RTL word and tag comparisons
- 35 / 35 functional points and 24 / 24 same-transaction crosses
- 16-vector back-to-back stream at 1.000 vectors/cycle
- 8 named assertions and 4 / 4 detected RTL mutations
- 88.24% raw and 100.00% reviewed executable RTL line coverage, with all 10 exclusions listed
- PyTorch nn.Linear quantization demo and Yosys synthesis proxy
The project is a compact 4x4 signed INT8 dot-product engine and does not claim a production NPU, commercial coverage closure, or physical-design signoff.