hn.today

GPGPU-SIM – Cycle-Level Simulator for Nvidia GPUs with CUDA or OpenCL Workloads

github.com3 points1 comments
Screenshot of GPGPU-SIM – Cycle-Level Simulator for Nvidia GPUs with CUDA or OpenCL Workloads

GPGPU-Sim is a cycle-level simulator that models contemporary NVIDIA GPUs executing CUDA and OpenCL workloads, with support for advanced features such as Tensor Cores, CUDA Dynamic Parallelism, and an integrated visualization tool (AerialVision). It includes a configurable, validated power model (AccelWattch, which replaced GPUWattch) and supports execution-driven simulation paths including CuDNN and PyTorch, checkpointing, and debugging aids for functional simulation errors. The simulator has been validated against real hardware, tested across many CUDA versions (from 4.2 up through 12), and is compatible with the Accel-Sim framework to run SASS trace-based simulations generated by NVIDIA’s NVBit dynamic binary instrumentation.

The distribution bundles source, configuration files, build scripts (CMake and Make), performance/power tools, and utilities such as cuobjdump_to_ptxplus and OpenCL/CUDA runtime libraries. The interconnect uses Booksim, PTX semantics use the CUDA math library, and AccelWattch leverages McPAT for energy estimation (validated on NVIDIA Volta QV100). The project includes detailed install/build/run instructions, changelogs, and citations for papers describing Accel-Sim, AccelWattch, Tensor Core modeling, and related tools. Development is community-driven with many academic contributors; support is provided on mailing lists and via a bug tracker, though usage does not imply formal support.

Read on github.com1 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Science

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.