- High Performance Machine Learning at the Edge

| May 12, 2020

Over the past decade machine learning has continued to revolutionize computer vision. Convolutional neural networks can now outperform humans on a wide range of image classification, object detection, object tracking, and semantic segmentation tasks.

Stated more plainly - AI is changing the game for computer vision.

As such, I have been spending time studying the computer vision market and trying to understand what opportunities exist up and down the stack. Today we are already seeing examples of improved performance for existing applications - such as inspecting manufacturing processes, security and surveillance, and robotics. We also see entirely new applications that were not possible in the past - such as semi-autonomous vehicles.

While they’ll be many exciting full-stack solutions for specific use cases, it’s clear that there is an opportunity for new technology to facilitate computer vision at the lower infrastructure layers, including chips. Model training requires massive parallelization and inference requires throughput optimization both beyond the likes of what has been required in the past.

Some AI compute markets appear to be increasingly competitive and well served. Fighting against NVIDIA, Intel, Google, and now Amazon in the datacenter seems like a tough place for a new startup to compete. Similarly, serving the needs of mass consumer electronics seemed equally daunting and perhaps even more challenging sector in which to extract serviceable margins.

However, the edge AI chip market has a number of attractive characteristics. It’s a new, fast-growing market. There isn’t really a deeply established incumbent known for ML/AI. Managing power, dissipation, and memory in the way they are constrained at the edge require a different frame of reference than companies who design datacenter chip architectures.


When Krishna first came to my office in June of 2019, it was clear he’d spent much of his career to date developing the skills necessary to found and lead a chip company. He rose up through the ranks of Xilinx from an engineer to run all aspects of their business as the GM for their $2.5B+ business and also as their EVP of sales.

He told me that he was going to become the computer vision partner of choice for customers focused on the embedded edge and that he has the team to do so. It’s hard to fully articulate just how exceptional the pool of talent he has assembled is. It includes technical engineering Fellows from both of the major chipmakers, some of the best chip and compiler architects on the planet, and our independent board member Moshe Gavrielov who also sits on the board of TSMC.

It took a few meetings together to come up with the best metric around which to frame’s value proposition. We eventually settled on frames per second per Watt (FPS/W). It’s a single measure that describes the efficiency of the solution when applied in the domain of computer vision. Simply put, delivers the highest FPS/W, offering greater than 30x improvement compared to other competing solutions.

Ironically, despite being an AI chip which one would expect to be an extraordinary hardware undertaking,’s MLSoC is a semiconductor-lite project in some ways. The intuition is pretty simple - every time you etch something in silicon it’s going to use power. If you want the most power-efficient architecture possible, you need to keep the silicon as simple and move as much of the smarts as possible into the compiler and software stack. Thus, in some sense, is more of a full-stack solution than meets the eye. It even works with existing models so you don’t have to redevelop algorithms specific to the MLSoC platform.

Wing Partnership

Founders Krishna Rangasayee and Steve Rosston to my left and right

All of us at Wing were thrilled to partner with the team at and lead their Power Seed last year. Wing invested behind Krishna’s vision to make the trusted computer vision partner for the embedded edge market.’s MLSoC solution offers groundbreaking energy efficiency that has the potential to greatly expand what is possible to do on the edge and to provide significant runtime improvements in power-constrained environments.

Today, we celebrate coming out of stealth and announcing their financings to date. We are excited to welcome Dell Technologies Capital into the fold, who is leading’s Series A. See TechCrunch coverage here.

Share this Article