MaxSim Scoring on L2-normalized FP16 embeddings Nq=32, d=128
0.049Latency (ms)PLAID GPU
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PLAID GPUNd=128, B=1K2026.06 | 0.049 | — | — | |
| TileMaxSim V2-MQNd=128, B=1K2026.06 | 0.054 | 0.9 | 1.5 | |
| torch.comp.Nd=128, B=1K2026.06 | 0.083 | — | — | |
| TileMaxSim V2-MQNd=128, B=10K2026.06 | 0.131 | 1.8 | 3.5 | |
| TileMaxSim V2-MQNd=256, B=10K2026.06 | 0.219 | 1.9 | 4.1 | |
| PLAID GPUNd=128, B=10K2026.06 | 0.233 | — | — | |
| PLAID GPUNd=256, B=10K2026.06 | 0.412 | — | — | |
| torch.comp.Nd=128, B=10K2026.06 | 0.458 | — | — | |
| TileMaxSim V2-MQNd=128, B=50K2026.06 | 0.594 | 1.8 | 3.8 | |
| torch.comp.Nd=256, B=10K2026.06 | 0.905 | — | — | |
| PLAID GPUNd=128, B=50K2026.06 | 1.045 | — | — | |
| TileMaxSim V2-MQNd=128, B=100K2026.06 | 1.195 | 1.7 | 3.7 | |
| PLAID GPUNd=128, B=100K2026.06 | 2.059 | — | — | |
| torch.comp.Nd=128, B=50K2026.06 | 2.235 | — | — | |
| torch.comp.Nd=128, B=100K2026.06 | 4.437 | — | — |