Reproduce “RedMule: A Mixed-Precision Matrix-Matrix Operation Engine for Flexible and Energy-Efficient On-Chip Linear Algebra and TinyML Training Acceleration” 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
Yvan Tortorella, Luca Bertaccini, Luca Benini, Davide Rossi, Francesco Conti
RedMulE, a specialized mixed-precision matrix-matrix operation engine, accelerates on-chip linear algebra and TinyML training by leveraging FP16 and hybrid FP8 formats. It achieves up to 117 GFLOPS at 920 GFLOPS/W for FP8 operations and provides a 6.4x speedup for MobileNet V2 training over software on edge devices.
Key insights
- A parametric micro-architecture with fine-grained hierarchical clock gating dynamically disables unused computational elements and buffer lines, achieving up to 37% power savings during matrix "leftovers" processing.
- A mixed-precision strategy, leveraging 8-bit inputs and 16-bit internal computation and output precision, provides negligible accuracy loss while significantly enhancing energy efficiency and performance.
- Extending Computing Elements with compact Fused Multiply-Add (FMA) and Floating-Point Non-Computational Operations (FNCOMP) units enables comprehensive GEMM-Ops support with only a 16% area overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.