Processor and method of operating a processor

Abstract

A processor and a method of operating a processor are disclosed. In some embodiments, the method includes forming a first set of products and forming a second set of products. The step of forming the first set of products may include: multiplying a first activation value with a first least significant sub-word, a second least significant sub-word, and a most significant sub-word in a first multiplier, a second multiplier, and a third multiplier; and adding the resulting first partial product and the resulting second partial product. The step of forming the second set of products may include forming a first floating point product, the step of forming the first floating point product including multiplying in a first multiplier a first sub-word of a mantissa of the activation value by a first sub-word of a mantissa of the weight to form a third partial product.

Description

[0001] This application claims SYSTEM AND METHOD FOR IMPROVING AREAANDPOWER EFFICIENCY BY REDISTRIBUTING WEIGHT NIBBLES AND SUPPORT FOR FP16)" U.S. Provisional Application No. 63 / 112,299 and U.S. Provisional Application No. 17 / 133,288, filed December 23, 2020, which are incorporated by reference in their entirety Included here. technical field [0002] One or more aspects of embodiments in accordance with the present disclosure relate to processing circuits, and more particularly, to processors and methods of operating processors. Background technique [0003] Processors used for neural networks can perform a large number of multiplication and addition operations, since a large portion of the numbers being processed may be relatively small, and only a small portion of outlier values ​​may be relatively large, so multiplication operations And some of the addition operations can be a poor use of processing resources. Furthermore, some operations in such a system may be integ...

AI Technical Summary

Problems solved by technology

[0003] Processors used for neural networks can perform a large number of multiplications and additions, because a large fraction of the numbers being processed may be relatively small, and only a small fraction of outlier values ​​may be relatively large, so the multiplication and some of the addition operations may be a poor use of processing resources

Furthermore, some operations in such systems may be integer operations and some operations may be floating point operations, which would consume significant chip area and power if performed on separate sets of dedicated hardware

Images