Matrix Multiplication Latency on Synthetic Matrix Multiplication Shapes
9.24Latency (µs)BWTA
Evaluation Results
| Method | Links | |
|---|---|---|
| BWTAShape=W: [768, 768], A: [128, 768]2026.04 | 9.24 | |
| BWTAShape=W: [3072, 768], A: [128, 768]2026.04 | 10.43 | |
| QTIP (r2)M=1, N=2048, K=20482025.12 | 19.44 | |
| QUIP# (e8p)M=1, N=2048, K=20482025.12 | 19.47 | |
| cuBLASM=8, N=2048, K=20482025.12 | 19.79 | |
| cuBLASM=1, N=2048, K=20482025.12 | 19.82 | |
| cuBLASM=4, N=2048, K=20482025.12 | 19.99 | |
| AQLM (2x8)M=1, N=2048, K=20482025.12 | 20.55 | |
| CodeGEMM (m1v4)M=1, N=2048, K=20482025.12 | 20.66 | |
| CodeGEMM (m2v8)M=1, N=2048, K=20482025.12 | 20.75 | |
| QUIP# (e8p)M=1, N=2048, K=81922025.12 | 23.44 | |
| CodeGEMM (m1v4)M=1, N=2048, K=81922025.12 | 23.81 | |
| QUIP# (e8p)M=1, N=4096, K=40962025.12 | 23.96 | |
| CodeGEMM (m2v8)M=1, N=2048, K=81922025.12 | 24.28 | |
| BWTAShape=W: [768, 3072], A: [128, 3072]2026.04 | 24.38 | |
| CodeGEMM (m2v8)M=1, N=4096, K=40962025.12 | 24.76 | |
| QTIP (r2)M=1, N=2048, K=81922025.12 | 24.9 | |
| CodeGEMM (m1v4)M=1, N=4096, K=40962025.12 | 24.97 | |
| QUIP# (e8p)M=1, N=8192, K=20482025.12 | 25.52 | |
| CodeGEMM (m2v8)M=1, N=8192, K=20482025.12 | 25.94 | |
| CodeGEMM (m1v4)M=1, N=8192, K=20482025.12 | 26.7 | |
| QTIP (r2)M=1, N=4096, K=40962025.12 | 26.74 | |
| QTIP (r2)M=1, N=8192, K=20482025.12 | 27.08 | |
| cuBLASM=1, N=2048, K=81922025.12 | 27.52 | |
| cuBLASM=1, N=4096, K=40962025.12 | 28 | |
| cuBLASM=8, N=4096, K=40962025.12 | 28.11 | |
| cuBLASM=4, N=4096, K=40962025.12 | 28.54 | |
| cuBLASM=8, N=2048, K=81922025.12 | 28.69 | |
| AQLM (2x8)M=1, N=8192, K=20482025.12 | 28.83 | |
| AQLM (1x16)M=1, N=2048, K=20482025.12 | 28.84 | |
| AQLM (1x16)M=1, N=8192, K=20482025.12 | 28.84 | |
| cuBLASM=4, N=2048, K=81922025.12 | 29.82 | |
| cuBLASM=1, N=8192, K=20482025.12 | 30.57 | |
| AQLM (2x8)M=1, N=2048, K=81922025.12 | 30.93 | |
| cuBLASM=4, N=8192, K=20482025.12 | 31.31 | |
| cuBLASM=8, N=8192, K=20482025.12 | 31.7 | |
| AQLM (2x8)M=1, N=4096, K=40962025.12 | 32.28 | |
| CodeGEMM (m1v4)M=1, N=4096, K=143362025.12 | 33.92 | |
| CodeGEMM (m1v4)M=1, N=8192, K=81922025.12 | 35.45 | |
| QTIP (r2)M=4, N=2048, K=20482025.12 | 36 | |
| CodeGEMM (m2v8)M=1, N=4096, K=143362025.12 | 36.15 | |
| QUIP# (e8p)M=4, N=2048, K=20482025.12 | 36.71 | |
| QUIP# (e8p)M=1, N=4096, K=143362025.12 | 37.27 | |
| CodeGEMM (m1v4)M=1, N=14336, K=40962025.12 | 37.51 | |
| CodeGEMM (m2v8)M=1, N=8192, K=81922025.12 | 37.99 | |
| QUIP# (e8p)M=1, N=8192, K=81922025.12 | 38.31 | |
| CodeGEMM (m2v8)M=1, N=14336, K=40962025.12 | 38.85 | |
| QUIP# (e8p)M=1, N=14336, K=40962025.12 | 38.91 | |
| cuBLASShape=W: [768, 768], A: [128, 768]2026.04 | 41.32 | |
| CodeGEMM (m1v4)M=4, N=2048, K=20482025.12 | 41.92 | |
| AQLM (2x8)M=4, N=2048, K=20482025.12 | 43.31 | |
| QTIP (r2)M=1, N=4096, K=143362025.12 | 43.85 | |
| CodeGEMM (m2v8)M=4, N=2048, K=20482025.12 | 44.04 | |
| QTIP (r2)M=1, N=8192, K=81922025.12 | 49.86 | |
| QTIP (r2)M=1, N=14336, K=40962025.12 | 51.3 | |
| QUIP# (e8p)M=4, N=2048, K=81922025.12 | 51.91 | |
| CodeGEMM (m1v4)M=4, N=2048, K=81922025.12 | 52.57 | |
| QUIP# (e8p)M=4, N=4096, K=40962025.12 | 53.92 | |
| CodeGEMM (m2v8)M=4, N=2048, K=81922025.12 | 56.21 | |
| CodeGEMM (m1v4)M=4, N=4096, K=40962025.12 | 57.79 | |
| QTIP (r2)M=8, N=2048, K=20482025.12 | 57.87 | |
| AQLM (2x8)M=1, N=4096, K=143362025.12 | 58.7 | |
| QTIP (r2)M=4, N=2048, K=81922025.12 | 59.03 | |
| QUIP# (e8p)M=8, N=2048, K=20482025.12 | 59.44 | |
| AQLM (1x16)M=1, N=2048, K=81922025.12 | 60.47 | |
| CodeGEMM (m2v8)M=4, N=4096, K=40962025.12 | 60.58 | |
| QUIP# (e8p)M=4, N=8192, K=20482025.12 | 60.7 | |
| AQLM (2x8)M=1, N=8192, K=81922025.12 | 62.5 | |
| QTIP (r2)M=4, N=4096, K=40962025.12 | 62.74 | |
| AQLM (1x16)M=1, N=4096, K=40962025.12 | 63.13 | |
| CodeGEMM (m2v8)M=4, N=8192, K=20482025.12 | 63.97 | |
| AQLM (2x8)M=1, N=14336, K=40962025.12 | 64.76 | |
| CodeGEMM (m1v4)M=4, N=8192, K=20482025.12 | 65.36 | |
| QTIP (r2)M=4, N=8192, K=20482025.12 | 66.18 | |
| CodeGEMM (m1v4)M=8, N=2048, K=20482025.12 | 69.72 | |
| AQLM (2x8)M=8, N=2048, K=20482025.12 | 73.03 | |
| AQLM (1x16)M=4, N=2048, K=20482025.12 | 74.67 | |
| AQLM (1x16)M=4, N=8192, K=20482025.12 | 74.67 | |
| CodeGEMM (m2v8)M=8, N=2048, K=20482025.12 | 75.18 | |
| AQLM (2x8)M=4, N=8192, K=20482025.12 | 76.15 | |
| CodeGEMM (m1v4)M=1, N=8192, K=286722025.12 | 76.5 | |
| CodeGEMM (m1v4)M=1, N=28672, K=81922025.12 | 76.71 | |
| cuBLASShape=W: [3072, 768], A: [128, 768]2026.04 | 80.17 | |
| AQLM (2x8)M=4, N=2048, K=81922025.12 | 82.18 | |
| CodeGEMM (m2v8)M=1, N=8192, K=286722025.12 | 86.2 | |
| cuBLASM=1, N=4096, K=143362025.12 | 86.31 | |
| CodeGEMM (m2v8)M=1, N=28672, K=81922025.12 | 86.48 | |
| cuBLASM=8, N=4096, K=143362025.12 | 86.49 | |
| cuBLASM=4, N=4096, K=143362025.12 | 86.51 | |
| cuBLASM=1, N=14336, K=40962025.12 | 88.67 | |
| cuBLASM=4, N=14336, K=40962025.12 | 89.08 | |
| cuBLASM=8, N=14336, K=40962025.12 | 89.29 | |
| AQLM (2x8)M=4, N=4096, K=40962025.12 | 89.76 | |
| QUIP# (e8p)M=8, N=2048, K=81922025.12 | 89.91 | |
| CodeGEMM (m1v4)M=8, N=2048, K=81922025.12 | 90.73 | |
| CodeGEMM (m1v4)M=4, N=4096, K=143362025.12 | 92.61 | |
| QUIP# (e8p)M=8, N=4096, K=40962025.12 | 93.43 | |
| cuBLASM=8, N=8192, K=81922025.12 | 95.45 | |
| cuBLASM=1, N=8192, K=81922025.12 | 96.4 | |
| CodeGEMM (m1v4)M=4, N=8192, K=81922025.12 | 98.66 |