Reproduce “Assessing Tenstorrent's RISC-V MatMul Acceleration Capabilities” 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
Hiari Pizzini Cavagna, Daniele Cesarini, Andrea Bartolini
Researchers from the University of Bologna and Cineca characterized the Tenstorrent Grayskull e75 RISC-V accelerator's matrix multiplication (MatMul) performance and energy efficiency against state-of-the-art CPUs and GPUs. The study revealed Grayskull's execution model, core scaling, and an energy efficiency reaching 1.56 TFLOPs/Watt, surpassing NVIDIA A100 at reduced precision.
Key insights
- Grayskull's execution model incurs significant compilation overhead during the first kernel run, while subsequent runs are dominated by data transfer when data is not persistent on the device.
- Performance scales effectively with an increasing number of active Tensix cores for larger matrices, showing up to 56x speedup when utilizing 64 cores.
- Optimized kernels leveraging sharded memory configurations and L1 SRAM significantly improve performance, especially when matrix dimensions fit within the L1 memory capacity, achieving 79.36% efficiency relative to theoretical peak.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.