GEMM Optimization
Performance numbers are teaching placeholders
The TFLOPS, speedup and memory figures on this page are reference values, not results reproduced by this repository on a pinned hardware/software stack. Re-measure on your own GPU before quoting them.
Detailed walkthrough of the 7-step GEMM optimization path.
Optimization Path
Summary
| Step | Technique | TFLOPS (FP32) | Speedup |
|---|---|---|---|
| 1 | Naive | 0.5 | 1.0× |
| 2 | Shared Memory | 2.0 | 4.0× |
| 3 | Double Buffer | 3.5 | 7.0× |
| 4 | Register Tiling | 6.0 | 12.0× |
| 5 | WMMA | 50+ | 100× |
| 6 | MMA PTX | 60+ | 120× |
| 7 | Pipeline | 70+ | 140× |