Reproduce “Decode Late, Multiply Once: Rethinking the Dot Product for Quantized Weights” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jonathan Tapson
The method reduces low-bit dot-product multiplications by grouping activations by weight code, then decoding and multiplying each group sum. It retains activation additions and requires code-based addressing. Integer arithmetic is exact, while floating-point summation can round differently. A projected fabric design sharing across 16 processing elements cuts LUTs per element from 819 to 179; this is not a measured complete-array or system-level gain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.