List your company

Matrix Multiplication on Synthetic Transformer Shapes Query-Key Q ⊗ K⊤

5.41Latency (µs)BWTA_QK
-107.7336655.98571,419.7052,183.4243Apr 5, 2026

Evaluation results

MethodLinks
BWTA_QKShape=Q: [128, 128], K: [128, 128]2026.045.41
BWTA_QKShape=Q: [512, 512], K: [512, 512]2026.048.96
FP16Shape=Q: [128, 128], K: [128, 128]2026.0439.31
cuBLASShape=Q: [128, 128], K: [128, 128]2026.0452.56
cuBLASShape=Q: [512, 512], K: [512, 512]2026.0458.23
BWTA_QKShape=Q: [2048, 2048], K: [2048, 2048]2026.04144.1
FP16Shape=Q: [512, 512], K: [512, 512]2026.04145.7
cuBLASShape=Q: [2048, 2048], K: [2048, 2048]2026.04606.5
FP16Shape=Q: [2048, 2048], K: [2048, 2048]2026.042,834