Search papers, labs, and topics across Lattice.
This paper introduces a cycle-level simulation framework that accurately models modern GPU architectures, addressing the limitations of existing simulation tools in capturing the physical non-uniformity and scale of contemporary AI workloads. The framework has been rigorously validated against physical silicon, achieving a 99% Pearson correlation coefficient and a 13.4% mean absolute cycle error on the H100 GPU. Through this simulation infrastructure, the authors explore architectural case studies that evaluate key design trajectories, such as chiplet topology scaling and inter-GPU prefetching strategies, which are critical for optimizing performance in the AI era.
Achieving a 99% correlation with real silicon, this simulation framework reveals crucial insights into the architectural evolution of GPUs for AI workloads.
The rapid evolution of machine learning workloads has fundamentally transformed GPU hardware, driving architectures toward Multi-Chip Module (MCM) topologies, asynchronous execution primitives, and persistent, multi-phase kernel behaviors. Despite these shifts, cycle-level simulation infrastructure has lagged behind, lacking the native capability to model the physical non-uniformity of modern GPUs alongside the massive scale of state-of-the-art AI workloads. To bridge this gap, we present a cycle-level simulation framework designed to accurately model modern GPU generations, including Ampere, Hopper, and Blackwell. Rigorously validated against physical silicon, the simulator achieves a 99% Pearson correlation coefficient and a 13.4% mean absolute cycle error on the H100 GPU. Utilizing this infrastructure, we conduct architectural case studies to evaluate emerging design trajectories, including chiplet topology scaling, expanded SRAM capacity and bandwidth, and inter-GPU prefetching strategies.
