accel-sim/accel-sim-framework

What it solves

Accel-Sim is a cycle-level GPU simulator designed to provide highly accurate performance modeling of NVIDIA GPUs. It solves the problem of needing real hardware for every architectural experiment or performance analysis, allowing developers and researchers to simulate contemporary workloads—including LLM inference and training—with high fidelity to real silicon.

How it works

The framework uses a trace-driven approach. It first uses an NVBit-based tracer to capture real SASS (Shader Assembly) execution from NVIDIA hardware. These traces are then replayed on a detailed performance model (GPGPU-Sim 4.x) combined with the AccelWattch power model.

Accel-Sim 2.0 specifically adds support for the NVIDIA Hopper architecture (H100/H200), modeling complex features like the Tensor Memory Accelerator (TMA), asynchronous Warp Group MMA (WGMMA), and threadblock clusters. For LLMs, it provides a PyTorch hook that allows users to trace specific representative layers of a model (e.g., within vLLM) rather than tracing the entire model, which would be computationally prohibitive.

Who it’s for

It is primarily for computer architects, GPU researchers, and AI engineers who need to analyze GPU hardware utilization, energy consumption, and the impact of architectural changes on AI workloads.

Highlights

• High Accuracy: Validated against 34,000+ kernel instances with a 99% Pearson correlation to real H100 silicon.
• Full Hopper Support: Models H100/H200 specific features including TMA, WGMMA, and mbarrier synchronization.
• LLM-Ready: Native support for tracing vLLM inference and PyTorch training steps without requiring hand-written kernels.
• Power Modeling: Integrated AccelWattch and GPUWattch models for energy analysis.
• Flexible Tracing: Supports both SASS trace-driven simulation and PTX execution-driven mode.