Blog

A few weeks ago, we announced the world’s fastest unified AI inference engine. The Modular AI Engine provides significant usability, portability, and performance gains for the leading AI frameworks — PyTorch and TensorFlow — and delivers world-leading execution performance for all cloud-available CPU architectures.

Last week, we launched Modular to the world after more than 16 months in stealth. We started Modular with a deep conviction — after 6+ years of building and scaling AI infrastructure to billions of users and 20+ years of building foundational compute infrastructure — it was clear the world needed a better path forward. Everyone wants less complexity, better access to compute and hardware, and the ability to develop and deploy AI faster.

We’re excited to finally share what we’ve been building at Modular. This announcement begins Modular’s journey to radically change the nature of AI programmability, usability, scalability, and compute.

In this post, we describe Modular’s approach to solving this problem and its game-changing benefits, including a new standard in state-of-the-art (SOTA) performance on CPU as compared to existing solutions.

AI is powered by a virtuous circle of data, algorithms (“models”), and compute. Growth in one pushes needs in the others and can grossly affect the developer experience on aspects like usability and performance. Today, we have more data and more AI model research than ever before, but compute isn’t scaling at the same speed due to … well, physics.

At Modular, we are rebuilding AI infrastructure for the world. Our goal is to move past AI tools that are themselves research projects and into a future where AI development and deployment are orders of magnitude more efficient for everyone. You should be able to do this without trading off performance or having to rewrite your entire code base.

The technological progress that has been made in AI over the last ten years is breathtaking — from AlexNet in 2012 to the recent release of ChatGPT, which has taken large foundational models and conversational AI to another level.

The first four requirements address one fundamental problem with how we've been using MLIR: weights are constant data, but shouldn't be managed like other MLIR attributes. Until now, we've been trying to place a square peg into a round hole, creating a lot of wasted space that's costing us development velocity (and, therefore, money for users of the tools).

Here in November 2022, we see a continuing onslaught of bad news: significant layoffs of incredible people as companies tighten their belts; companies that raised too much money, too fast, without core fundamentals are dying; and a changing climate where over-tightening rather than under-tightening is seemingly the new normal.

What makes a brand? When you see an apple with a bite out of it, you immediately associate it with the famous technology company — but what about the logo elicits the feelings it does?