Accel-Sim Framework Release Notes
Major Features and Improvements
Trace Generation and Processing
- Updated tracer tool to use NVBit 1.7, improving compatibility and performance [8d0785d]
- Added support for compressed (xz) trace file generation by default [b5f038e]
- Improved trace parser performance and memory usage [4be9303, 08b9e9c]
- Replaced std::cin with a custom PipeReader class for improved trace parsing [16b55f1]
- Supports execution from compressed traces, saving runtime disk space by ~20-30x.
Multi-Stream Support
- Added per-stream statistics collection and reporting [f783a4c]
- Implemented passing of CUDA stream IDs from Accel-Sim to GPGPU-Sim for accurate multi-stream simulation [f783a4c]
Hardware Correlation
- Updated correlation scripts to support parsing data from newer NVIDIA profiling tools [56c5950]
- Added support for correlating with NVIDIA H100 GPU data [3a8112f]
- Improved handling of hardware data collection for various GPU architectures [56c5950, 3a8112f]
Configuration and Benchmark Management
- Updated CUTLASS benchmark naming and configuration [8253fd6]
- Implemented truncation of excessively long simulation names to improve file system compatibility [76dfd0b]
CI/CD and Testing
- Updated CI workflows to use the latest hardware data and improve correlation testing [56c5950, 01feb3c]
- Implemented automated code formatting checks in the CI pipeline [4a90ea7]
Bug Fixes and Optimizations
- Fixed issues with running hardware traces, including proper handling of kernel limits [49fe436]
- Resolved compatibility issues with newer CUDA versions and GPU architectures [8d0785d]
- Improved error handling and reporting in various scripts [3a8112f, 49fe436]
Documentation and Usability
- Updated README files and documentation to reflect recent changes [b5f038e]
- Improved error messages and user guidance for environment setup [8d0785d]
Acknowledgments
This release represents the collective effort of numerous contributors, including researchers and developers from various institutions. Their dedication has significantly advanced the capabilities of GPU architectural simulation.