Fast and accurate models in computer-vision are critical for optimizing costs in cloud services and enabling compute on edge-devices. Often papers & libraries report GFLOPs & model size as a proxy metric for speed, but is that reliable?

Introduction

In today’s fast-paced era, where development frequently pivots and agile processes reign supreme, the notion of a Technical Design Document (TDD) might seem antiquated to some. Why invest time in meticulously documenting something that might change tomorrow? But ask any technology leader and they will remind you of the adage “failing to plan is planning to fail”, which remains resoundingly true.

Navigating the intricate realm of software development is akin to constructing a skyscraper. Just as an architect relies on blueprints to shape towering structures, software projects lean on Technical Requirements Documents (TRDs) to achieve vision and purpose. They provide a concise and objective perspective that is far broader than functional User Stories alone.

In today’s competitive landscape, the ability to create robust, scalable, and efficient systems is no longer a luxury - it’s a necessity. From tech giants like Amazon & Google to emerging start-ups, the common thread among successful organizations is an effective system design process.

valarray is probably the least known standard library container, it uses expression templates to optimize compute optimizations. This technique is the foundation of most numerical libraries such as Eigen, PyTorch, and TensorFlow.