Precision on demand : a novel lossles mixed-precision computation technique

Abstract EN

Nowadays, there are wide range of computing-intensive applications that require a powerful computing platform.

This computational complexity can be reduced significantly using lower precision, but certainly the accuracy will be affected.

In this work, a novel lossless mixed-precision computation technique is used to reduce the computational complexity for such problems while keeping the same accuracy of higher precision.

The “precision on demand” technique depends on iterative computation method, which utilizes discarding stages.

The discarding stages are commonly used in many applications where some intermediate results are discarded, and they do not contribute to the final results.

Max, min, threshold, and ReLU -operations are examples of such discarding stages.

Lower precision is used to compute all intermediate results, then non-discarded values are recomputed using higher precisions.

This technique enhances performance, improves power-consumption, reduces memory requirements, and allows implementing heavy computational systems on low resources and low-cost platforms.

This work presents the decision on demand concept, and discuss many details related to hardware architecture implementations and optimizations.

CNN inference is used as a case study, and speed is improved by a factor of 1.6x- 3.5x. 

ی