Complexity and Throughput Analysis on Hardware Efficiency Benchmark H100 Configuration
93.8H100 Throughput (tokens/s)GDN-2 (matrix-gated)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GDN-2 (matrix-gated)Hardware=Single H100, Model Scale=125M, Sequence length (T)=1024, Mini-batch (mb)=82026.06 | 93.8 | — | — | — | — | |
| CARVEHardware=Single H100, Model Scale=125M, Sequence length (T)=1024, Mini-batch (mb)=82026.06 | 93.4 | — | — | — | — | |
| Content-aware exact recurrentHardware=Single H100, Model Scale=125M, Sequence length (T)=1024, Mini-batch (mb)=82026.06 | 24.8 | — | — | — | — | |
| Delta RuleHardware=Single H100, Model Scale=125M, Sequence length (T)=1024, Mini-batch (mb)=82026.06 | — | 2 | 2 | 2 | — | |
| Linear AttentionHardware=Single H100, Model Scale=125M, Sequence length (T)=1024, Mini-batch (mb)=82026.06 | — | 2 | 2 | 2 | — |