Reproduce “A Lightweight High-Throughput Collective-Capable NoC for Large-Scale ML 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, Lorenzo Leone, Chen Wu, Tim Fischer, Raphael Roth, Luca Benini
The system adds in-network multicast and reduction to a tiled accelerator, using Direct Compute Access to borrow tile compute units for arithmetic reductions. Cycle-accurate simulations on a 4×4 mesh report geomean speedups of 5.3× for multicast and 2.8× for reduction over optimized software baselines for 1–32 KiB transfers. GEMM gains are modeled, not end-to-end inference measurements.
Key insights
- High-performance, scalable on-chip collective communication is essential for overcoming data movement bottlenecks in next-generation manycore ML accelerators.
- Integrating in-network computation via Direct Compute Access (DCA) allows for efficient, low-overhead hardware acceleration of complex operations like arithmetic reduction by leveraging existing FPU resources.
- Efficient on-chip hardware support for many-to-one traffic is feasible and can significantly improve the scalability and energy efficiency of ML accelerators by reducing communication overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.