A dynamic block floating-point format with pivot-focus and adaptive grouping, plus a hierarchical lookup table, lets attention Softmax run in integer-only hardware with negligible accuracy loss.
Reduced Softmax Unit for Deep Neural Network Accelerators
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
abstract
The Softmax activation layer is a very popular Deep Neural Network (DNN) component when dealing with multi-class prediction problems. However, in DNN accelerator implementations it creates additional complexities due to the need for computation of the exponential for each of its inputs. In this brief we propose a simplified version of the activation unit for accelerators, where only a comparator unit produces the classification result, by choosing the maximum among its inputs. Due to the nature of the activation function, we show that this result is always identical to the classification produced by the Softmax layer.
fields
cs.AR 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Pushing the Limits of BFP on Narrow Precision LLM Inference
A dynamic block floating-point format with pivot-focus and adaptive grouping, plus a hierarchical lookup table, lets attention Softmax run in integer-only hardware with negligible accuracy loss.