Benchmark Explorer

Interactive benchmark results from NVIDIA T4 (Kaggle). Entries marked [P] are projected from single-device measurements with 95% linear scaling efficiency.

Attention Latency (ms) vs Sequence Length

Latency Data

seq_len	Dense (ms)	HADS (ms)	Reduction	BigBird (ms)	Sliding (ms)
512	12.4	8.1	34.7%	5.6	3.8
1024	47.3	28.9	38.9%	19.4	11.2
2048	181.2	105.6	41.7%	84.2	64.3
4096	713.8	412.4	42.2%	298.3	219.7