Ilmu Komputer & AI editorial
Open AccessOA2026
FlashGPU-sim: Enabling GPU Modeling for Modern Architectures and AI Workloads
An open-source, execution-driven, cycle-accurate simulator for modern AI workloads, achieving 5.24% MAPE across 131 configurations on RTX 5090, H100, and B200.
Siying Yu; Yixun Hong; Guozhi Qiu; Jingci Liu; Feng Gu; Chenbo Geng; Zhengrong Wang; Chen Zhang; Bei Yu· 2026· DOI 10.48550/arXiv.2609.15311
The core problem
The rapid advancement of artificial intelligence has created a tight software-hardware co-design loop, where modern GPUs expose features such as asynchronous data movement, tensor core pipelines, and fine-grained synchronization that high-performance kernels aggressively exploit. Simultaneously, emerging application behaviors increasingly influence the next generation of hardware design. However, the latest open-source simulators for NVIDIA GPUs focus on architectures and software stacks from roughly six years ago. Consequently, they cannot support many state-of-the-art AI kernels generated by modern compiler stacks, such as Triton, or accurately model the hardware features they depend on. As a result, architects lack a credible platform for analyzing bottlenecks in this flywheel or evaluating design trade-offs for future AI systems. To bridge this gap, the authors present FlashGPU-sim, an open-source, execution-driven, cycle-accurate GPU simulator for modern AI workloads. FlashGPU-sim faithfully models modern hardware features such as asynchronous data movement, fine-grained synchronization, tensor-core execution, and distributed shared memory. A Triton extraction front-end allows
Innovation
FlashGPU-sim was evaluated across 131 workload configurations on three modern GPU architectures: RTX 5090, H100, and B200. The key results are:
- **Cycle-level accuracy**: The simulator achieves a MAPE of 5.24%, indicating high fidelity in modeling cycle counts compared to real hardware.
- **Multi-threaded performance**: With 16 host threads, FlashGPU-sim achieves a 7.86x speedup over single-threaded simulation, making large-scale design space exploration feasible.
- **Case study**: An H100 case study demonstrates the simulator's utility for microarchitectural design exploration, enabling architects to analyze bottlenecks and evaluate trade-offs.
The following Mermaid diagram illustrates the simulation flow:
These results confirm that FlashGPU-sim can accurately and efficiently simulate modern AI workloads, bridging the gap left by older simulators.
The rapid advancement of artificial intelligence has created a tight software-hardware co-design loop, where modern GPUs expose features such as asynchronous data movement, tensor core pipelines, and fine-grained synchronization that high-performance kernels aggressively exploit. Simultaneously, emerging application behaviors increasingly influence the next generation of hardware design. However, the latest open-source simulators for NVIDIA GPUs focus on architectures and software stacks from roughly six years ago. Consequently, they cannot support many state-of-the-art AI kernels generated by modern compiler stacks, such as Triton, or accurately model the hardware features they depend on. As a result, architects lack a credible platform for analyzing bottlenecks in this flywheel or evaluating design trade-offs for future AI systems. To bridge this gap, the authors present FlashGPU-sim, an open-source, execution-driven, cycle-accurate GPU simulator for modern AI workloads. FlashGPU-sim faithfully models modern hardware features such as asynchronous data movement, fine-grained synchronization, tensor-core execution, and distributed shared memory. A Triton extraction front-end allows direct simulation of optimized AI operators without manual porting, while multi-threaded execution makes large-scale software-hardware co-design practical. Across 131 workload configurations on RTX 5090, H100, and B200, FlashGPU-sim achieves a cycle-level MAPE of 5.24%, while multi-threaded simulation reaches a 7.86x speedup with 16 host threads. An H100 case study further demonstrates its utility for microarchitectural design exploration.
FlashGPU-sim is designed as an execution-driven, cycle-accurate simulator that models key modern GPU hardware features. The simulator incorporates:
Why it matters
The introduction of FlashGPU-sim addresses a critical need in the AI systems community: a credible platform for analyzing bottlenecks in the software-hardware co-design loop and evaluating design trade-offs for future AI systems. By supporting modern compiler stacks like Triton and modeling contemporary hardware features, it enables architects to study state-of-the-art AI kernels without manual porting. The 5.24% MAPE demonstrates that the simulator is cycle-accurate enough for microarchitectural exploration, while the 7.86x speedup with 16 threads makes it practical for large-scale studies. The H100 case study further validates its utility. However, limitations may include the scope of supported architectures and the overhead of the Triton front-end. Future work could extend support to more GPU generations and additional compiler stacks. Overall, FlashGPU-sim represents a significant step forward in enabling rigorous, open-source GPU simulation for modern AI workloads, fostering innovation in both hardware and software design.
Who should read this
CS practitioners and researchers
Opening member content…