Search papers, labs, and topics across Lattice.
This paper evaluates the performance of three parallelization strategies for gravitational N-body simulations on the Tenstorrent Wormhole, a RISC-V-based accelerator. The study compares the execution time and energy consumption of each strategy to identify the optimal configuration for this scientific computing workload. Results highlight the most efficient approach for scaling N-body simulations on this novel architecture, balancing performance and energy efficiency.
RISC-V accelerators, originally for AI, can efficiently run scientific simulations, but only with the right parallelization strategy.
While RISC-V-based accelerators were initially designed with artificial intelligence applications in mind, they are increasingly being recognized as promising platforms for high performance scientific computing. In this work, we present three strategies for scaling an $N$-body code across multiple Tenstorrent Wormhole accelerators based on the RISC-V architecture. We assess the performance of these approaches by measuring both the execution time and the energy consumption required to complete a representative simulation, ultimately identifying the configuration that offers the most favorable balance between efficiency and performance.
