Sudheendra Kadri

Title: Sudheendra Kadri: Innovator in Deep Neural Network Memory Layouts.

Introduction

Sudheendra Kadri is an inventor based in Folsom, California, and works at Intel Corporation. His patent work addresses how data is organized and transferred for deep neural network computations.

Latest Patents

Kadri is associated with 1 patent, titled “Transposing Memory Layout of Weights in Deep Neural Networks (DNNs).” The invention describes a compute block with a direct memory access (DMA) engine that transfers data from external memory to local memory, where a multiply-accumulate (MAC) array can use it for convolution operations.

The patent describes storing filter weights in sections organized by channel, then converting them into a layout in which each new section contains a vector of weights drawn from different channels. The DMA engine may also compress the weights by removing zero-valued entries before converting the layout. This approach focuses on preparing neural network data for efficient processing within the compute block.

Career Highlights

Kadri’s disclosed patent work connects memory organization, data movement, and neural network computation. His association with Intel Corporation places this work in the context of the company’s computing and semiconductor technology activities.

Collaborations

Kevin Brady is identified as Kadri’s coworker. The available information does not specify whether Brady collaborated on the patent.

Conclusion

Sudheendra Kadri’s patent describes a method for reorganizing and optionally compressing neural network weights as they move into a compute block, highlighting the role of data layout in DNN processing.