List your company

Long-Context LLM Inference Prefill Performance

0.62Prefill Latency (ms)Kascade
-87.0508504.72711,096.5051,688.2829Dec 18, 2025

Evaluation results

MethodLinks
KascadeSeqlen=8192, Topk%=10%, Batch size=12025.120.62—1.231.62
FA3Seqlen=8192, Topk%=10%, Batch size=12025.120.76———
Tilelang (TL)Seqlen=8192, Topk%=10%, Batch size=12025.121———
KascadeSeqlen=262144, Topk%=10%, Batch size=12025.12408.3—2.122.57
KascadeSeqlen=262144, Topk%=30%, Batch size=12025.12727.55—1.191.44
AnchorSeqlen=262144, Topk%=30%, Batch size=12025.122,192.392.09——