LLM Decoding Throughput on NOSA-8B 32K context length
2,312.1Throughput (tok/s)SparDA
Evaluation Results
| Method | Links | |
|---|---|---|
| SparDABatch size=128, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 2,312.1 | |
| SparseBatch size=128, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 2,145.5 | |
| SparDABatch size=64, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 2,106.6 | |
| SparseBatch size=64, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 1,861.6 | |
| SparDABatch size=32, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 1,579.9 | |
| SparseBatch size=32, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 1,244.9 | |
| SparseBatch size=16, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 947.1 | |
| SparDABatch size=16, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 873.2 | |
| SparseBatch size=16, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 795.8 | |
| SparseBatch size=8, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 488.1 | |
| SparseBatch size=8, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 478 | |
| DenseBatch size=16, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 439.7 | |
| SparDABatch size=8, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 433.2 | |
| InfiniGenBatch size=128, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 356.5 | |
| DenseBatch size=8, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 350.7 | |
| InfiniGenBatch size=64, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 269.9 | |
| DenseBatch size=4, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 252 | |
| SparseBatch size=4, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 244 | |
| SparseBatch size=4, Offloading=Disabled, Hardware=NVIDIA H1002026.06 | 241.3 | |
| SparDABatch size=4, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 214.7 | |
| InfiniGenBatch size=32, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 205.9 | |
| InfiniGenBatch size=16, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 166 | |
| InfiniGenBatch size=8, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 120.3 | |
| InfiniGenBatch size=4, Offloading=Enabled, Hardware=NVIDIA H1002026.06 | 82 |