ResearchBenchmarksSparse Attention Acceleration on Transformer Sparse Attention WorkloadSparse Attention Acceleration on Transformer Sparse Attention WorkloadFollow0.1Accuracy Loss (Sparse Attn)Sanger0.0760.2380.40.562Dec 2, 2025Evaluation resultsMethodMethodLinksAccuracy Loss (Sparse Attn)AreaPowerAttention ThroughputNormalized Energy EfficiencyNormalized Area EfficiencySangerTech=55 nm, Freq=500 MHzTech=55 nm, Freq=500 MHz2025.120.116.92.762,1162,9581,025ESACTTech=28 nm, Freq=500 MHzTech=28 nm, Freq=500 MHz2025.120.25.090.7925,2886,6771,039SpAttenTech=40 nm, Freq=1 GHzTech=40 nm, Freq=1 GHz2025.120.71.550.3253602,261677