ConceptSSA
Subquadratic Sparse Attention
A sparse attention mechanism in which each token attends only to a selected subset of key tokens, reducing the computational complexity of attention from quadratic O(N²) to approximately linear O(N). This allows models to process extremely long contexts efficiently.
Timeline (last 90 days)
Oct 13
Subquadratic claims that SubQ Snap's Subquadratic Sparse Attention (SSA) can reduce attention cost from quadratic to linear, returning 100K tokens within 1 second, but these figures have no hardware specification or third-party verification.
Unverified50%