TVM Matrix Multiplication Optimization - Step 7: cuBLAS Comparison
Published:
Step 7: cuBLAS Comparison
Results
Overall Comparison (Average)
| Implementation | Performance | vs cuBLAS |
|---|---|---|
| NumPy (CPU) | 13 GFLOPS | 0.6% |
| TVM Step 6 | 1039 GFLOPS | 50.1% |
| cuBLAS (NVIDIA) | 2074 GFLOPS | 100% |
Size-Specific Details
| Size | TVM Step 6 (Unrolling) | cuBLAS | TVM/cuBLAS |
|---|---|---|---|
| 512x512 | 1028 GFLOPS | 1302 GFLOPS | ~79.0% |
| 1024x1024 | 1050 GFLOPS | 2846 GFLOPS | ~36.9% |
Analysis
Performance Characteristics of TVM Step 6 (Unrolling)
Around 79% at 512x512 size:
- For small matrices, the combination of Tiling, Shared Memory, Software Pipelining, and Loop Unrolling allows TVM to reach a high fraction of cuBLAS performance.
Around 37% at 1024x1024 size:
- For larger matrices, cuBLAS’s advanced optimizations (e.g., Tensor Core utilization, more aggressive tiling/vectorization) remain more effective.
Execution
# cuBLAS benchmark
python benchmarks/cublas_baseline.py
# TVM vs cuBLAS comparison
python benchmarks/compare_all_with_cublas.py
Code can be found at https://github.com/kimm240/matrix-multiplication-optimization-with-tvm.
Series Posts
- Previous: Step 6: Loop Unrolling
Language: 한국어 (Korean)