Grouped Query Attention on T4 GPU Synthetic Performance Benchmark
20.12Performance (TFLOPS)Flex Attention
Evaluation Results
| Method | Links | |
|---|---|---|
| Flex AttentionHead Dimension=64, Sequence Length=4k2025.06 | 20.12 | |
| DeepSeek-V3 + OursHead Dimension=128, Sequence Length=2k2025.06 | 18.82 |