Loading patent…
AssigneeNext Silicon Ltd
InventorsOded Margalit, Ilan Tayari
Estimating the term…
A digital hardware circuit and method for performing reduction operations that achieves constant timing depth regardless of input count. The invention replaces conventional sequential binary tree approaches with a parallel matrix comparison architecture where multiple comparator circuits simultaneously compare input signals against each other. Combinatorial logic circuits process comparison outputs to generate dominance signals indicating which input satisfies the reduction criteria, and selection logic outputs the final result. The parallel approach maintains exactly three logic levels regardless of number of processing inputs, enabling significantly higher clock frequencies than conventional methods whose timing depth increases logarithmically with input count. Applications include matrix multiplication engines, floating-point arithmetic units, and artificial intelligence accelerators where reduction operations for maximum/minimum finding, normalization, and other computations are performed millions of times per second. The constant timing depth enables operation at frequencies exceeding 1 GHz while providing scalable performance.
A method of operating a digital hardware circuit to find a reduction operation value among a plurality of input signals, the method comprising:
A digital hardware circuit for finding a reduction operation value among a plurality of input signals, comprising:
A method of operating a digital hardware circuit comprising:
Dependencies are inferred from the claim text.