Matrix Multiplication on Synthetic Transformer Shapes Query-Key Q ⊗ K⊤
5.41Latency (µs)BWTA_QK
Evaluation Results
| Method | Links | |
|---|---|---|
| BWTA_QKShape=Q: [128, 128], K: [128, 128]2026.04 | 5.41 | |
| BWTA_QKShape=Q: [512, 512], K: [512, 512]2026.04 | 8.96 | |
| FP16Shape=Q: [128, 128], K: [128, 128]2026.04 | 39.31 | |
| cuBLASShape=Q: [128, 128], K: [128, 128]2026.04 | 52.56 | |
| cuBLASShape=Q: [512, 512], K: [512, 512]2026.04 | 58.23 | |
| BWTA_QKShape=Q: [2048, 2048], K: [2048, 2048]2026.04 | 144.1 | |
| FP16Shape=Q: [512, 512], K: [512, 512]2026.04 | 145.7 | |
| cuBLASShape=Q: [2048, 2048], K: [2048, 2048]2026.04 | 606.5 | |
| FP16Shape=Q: [2048, 2048], K: [2048, 2048]2026.04 | 2,834 |