Reproduce “A Multicast-Capable AXI Crossbar for Many-core Machine Learning Accelerators” 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
Luca Colagrande, Luca Benini
Researchers at ETH Zurich developed a multicast-capable AXI crossbar for many-core machine learning accelerators that use non-coherent, scratch-pad memory. This hardware-supported multicast significantly improved data transfer efficiency, demonstrating up to a 16.2x speedup in DMA microbenchmarks and a 3.4x performance increase for matrix multiplication over a unicast baseline, with only modest area and timing overheads.
Key insights
- Hardware-level multicast support is crucial for overcoming memory bandwidth limitations in many-core ML accelerators by significantly reducing redundant data fetches for common broadcast patterns.
- A scalable mask-form encoding for AXI's `aw_user` signal can effectively enable multicast communication without requiring fundamental changes to the AXI protocol or complex address decoding for massive parallelism.
- Achieving substantial performance gains, such as a 16.5x higher operational intensity in matrix multiplication, through hardware multicast can be realized with a modest hardware overhead for the crossbar.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.