GTSim v0.1.0

GTSim (GPU-Tile-Sim) is a tile-centric GPU simulation framework for LLM hardware-software co-design. It executes a warp-centric tile graph whose nodes describe tile-level operations and whose edges encode data and order dependencies.

This repository accompanies GPU-Tile-Sim: A Tile-Centric GPU Simulation Framework for LLM Hardware-Software Co-Design. This paper is accepted and to appear in the proceedings of MICRO 2026.

This public core release provides the graph-driven simulator core and two H100 reference workload builders:

• a persistent two-stage H100 GEMM; and
• an H100 FA3 two-stage pipeline.

It is a scoped, runnable artifact rather than a complete release of every frontend, workload, runtime optimization, or case-study feature described in the GTSim paper. See the v0.1.0 release scope before citing or extending the repository.

Build and run

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

# H100 persistent two-stage GEMM: M N K
./build/gtsim_gemm 512 512 512

# H100 FA3 reference: batch heads_q heads_kv sequence_q sequence_kv
./build/gtsim_fa3 1 32 8 128 128

Run the smoke tests with ctest --test-dir build --output-on-failure. The checked reference outputs are in configs/v0.1.0/references.json. To re-run and compare those outputs, use:

python3 scripts/run_reference.py --build-dir build --verify

Run the measured H100 validation suite (8 GEMM and 6 FA3 cases) with:

python3 scripts/run_validation.py --build-dir build

It reports MAPE and Pearson correlation and writes the per-case report to validation/results/. The v0.1.0 result snapshot is in validation/v0.1.0.md.

License

The repository is distributed under the MIT license.