accel-sim/accel-sim-framework

This is the top-level repository for the Accel-Sim framework.

Accel‑Sim 2.0 – Cycle‑level GPU simulation for modern AI workloads

What it is – Accel‑Sim is an open‑source framework that lets researchers trace real GPU execution (SASS assembly) on NVIDIA hardware and then re‑play those traces on a detailed, cycle‑accurate simulator (GPGPU‑Sim 4.x) together with the AccelWattch power model. Version 2.0 adds full support for the latest Hopper GPUs (H100/H200), including new instructions, memory‑subsystem features, and a “GPUVision” profiling front‑end.

Why it matters for AI – The simulator can model LLM inference and training pipelines (vLLM, PyTorch) end‑to‑end without hand‑written kernels. The authors report 99 % Pearson correlation with real H100 hardware across >34 k kernels, giving confidence that performance and power estimates are realistic for large‑scale transformer models.

Key components

  • Tracer – NVBit‑based instrumentation that records SASS execution on a real GPU. It now supports compressed .tracez files and per‑layer hooks for PyTorch/vLLM.
  • SASS Front‑end / Simulation Engine – GPGPU‑Sim 4.x plus the Hopper‑specific micro‑architectural models (TMA, WGMMA, mbarrier, thread‑block clusters, chiplet‑aware L2, HBM3/e timing).
  • Correlator (GPUVision) – Tools to compare simulated cycles, instruction counts, and memory statistics against Nsight Compute data.
  • Tuner – Scripts for exploring configuration parameters and calibrating the model.
  • AccelWattch – Integrated power‑model that estimates energy consumption alongside performance.

Typical workflow

  1. Build the simulator (CMake + Python dependencies).
  2. Trace an application on a real GPU using the provided NVBit scripts (or the torch‑hook for LLM layers).
  3. Post‑process raw traces into the compressed .tracez format.
  4. Run the simulator with a Hopper config (H100‑SASS, H200‑SASS, etc.) via the run_simulations.py launch manager.
  5. Analyze results with built‑in monitoring, get_stats.py, and optional plotting scripts.

Getting started quickly – The repo ships a Docker image based on nvidia/cuda:12.8.0‑cudnn‑devel‑ubuntu24.04. After pulling the image you can follow the “Quick Start” section to build the simulator, trace a benchmark from the companion gpu‑app‑collection, and run a simulation in a few commands.

Pre‑collected traces – For users without access to an NVIDIA GPU, the release includes ready‑made Hopper traces for popular models (BERT, ResNet‑50, SDXL, Llama 3.1 8B, etc.). Download with get-accel-sim-traces.py, then simulate directly.

Roadmap – Upcoming work includes experimental support for the next‑generation Blackwell GPUs and multi‑GPU/NVLink configurations.

Who should use it – GPU architects, performance researchers, and AI system designers who need accurate, validated performance or energy estimates for new Hopper features or large‑scale LLM workloads.

Related

  • Project
    casys-kaist/LLMServingSim

    A cycle-level simulator for LLM inference infrastructure that models heterogeneous hardware, disaggregated memory, and advanced serving techniques to predict performance before deployment.

  • Project
    microsoft/vidur

    A high-fidelity LLM inference system simulator used to study performance, plan capacity, and test scheduling optimizations without requiring continuous GPU access.

  • Project
    isaac-sim/IsaacLab

    Isaac Lab is a GPU-accelerated, open-source robotics simulation framework built on NVIDIA Isaac Sim, designed to simplify and scale robot learning tasks like reinforcement learning and motion planning.

  • Dispatch
    Building a High-Performance LLM Inference Engine from Scratch with tiny-vLLM

    An exploration of tiny-vLLM, a C++ and CUDA-based educational project that teaches the internals of LLM inference, from Safetensors loading to PagedAttention.