### init: stable / GEMM time

| backend | k3-d1000 | k17-d180 | k100-d32 | k550-d6 | k3000-d2 |
| --- | ---: | ---: | ---: | ---: | ---: |
| Torch CPU | 1.03x | 0.99x | 0.99x | 1.29x | 1.51x |
| Torch CUDA | 1.26x | 1.19x | 1.19x | 1.21x | 1.53x |
| CuPy CUDA | 1.35x | 1.10x | 1.09x | 1.14x | 1.49x |

### lloyd: stable / GEMM time

| backend | k3-d1000 | k17-d180 | k100-d32 | k550-d6 | k3000-d2 |
| --- | ---: | ---: | ---: | ---: | ---: |
| Torch CPU | 1.14x | 1.14x | 1.28x | 1.77x | 1.39x |
| Torch CUDA | 1.94x | 1.70x | 2.26x | 2.74x | 2.73x |
| CuPy CUDA | 1.76x | 1.54x | 2.03x | 3.09x | 3.20x |

### fit: stable / GEMM time

| backend | k3-d1000 | k17-d180 | k100-d32 | k550-d6 | k3000-d2 |
| --- | ---: | ---: | ---: | ---: | ---: |
| Torch CPU | 1.11x | 1.12x | 1.18x | 1.40x | 1.46x |
| Torch CUDA | 1.80x | 1.47x | 1.55x | 1.73x | 1.91x |
| CuPy CUDA | 1.71x | 1.39x | 1.37x | 1.58x | 1.77x |

### Stable recomputation rate

| phase and backend | k3-d1000 | k17-d180 | k100-d32 | k550-d6 | k3000-d2 |
| --- | ---: | ---: | ---: | ---: | ---: |
| init, Torch CPU | 0.00011% | 0.00005% | 0.00007% | 0.00031% | 0.50631% |
| init, Torch CUDA | 0.00011% | 0.00005% | 0.00007% | 0.00031% | 0.50453% |
| init, CuPy CUDA | 0.00011% | 0.00005% | 0.00007% | 0.00031% | 0.50654% |
| lloyd, Torch CPU | 0.00001% | 0.00000% | 0.00001% | 0.00019% | 0.53142% |
| lloyd, Torch CUDA | 0.00001% | 0.00000% | 0.00001% | 0.00019% | 0.53142% |
| lloyd, CuPy CUDA | 0.00001% | 0.00000% | 0.00001% | 0.00019% | 0.53142% |
| fit, Torch CPU | 0.00002% | 0.00001% | 0.00002% | 0.00027% | 0.50684% |
| fit, Torch CUDA | 0.00002% | 0.00001% | 0.00002% | 0.00027% | 0.50528% |
| fit, CuPy CUDA | 0.00002% | 0.00001% | 0.00002% | 0.00027% | 0.50609% |
