Taking powerful artificial intelligence out of the realm of Clouds and into Chips
The conventional wisdom in today’s chip startups is a bit depressing: Raise a lot of capital, achieve diminishing improvements over what’s being offered by the ever-consolidating incumbents, compete against said incumbents.
At my firm, Lux Capital, we invest in all manner of futuristic machines, from drones at sea and in the air, to micro satellites, self-driving cars and rockets. We are constantly on the lookout for startups that are making radical improvements to the underlying technologies that make those machines go.
One area I was particularly driven to find: new chips that could empower tomorrow’s machines. Specifically: a chip that had a better way of shepherding electrons or photons towards training and running artificial intelligence at the edge.
We found it in Mythic over two years ago, when we invested in the company’s first round of venture capital funding.
Today Mythic is announcing a $40 million series B round led by SoftBank Ventures, with participation from existing investors, including Lux Capital, Data Collective, DFJ, AME Cloud Ventures, and new investors Andy Bechtolsheim and Lockheed Martin.
Conventional design techniques have made it very difficult to squeeze out extra performance. Moore’s Law is starting to deliver diminishing returns. As such, chips went from being the darling of Silicon Valley investors at the turn of the century to a four-letter word by the end of its first decade.
Rarely does a new chip company offer improvements in orders of magnitudes by rethinking how compute is done.
Every once in a while, a new type of chip enables an entirely new generation of machines. Mythic claimed to have invented an alternative to conventional logic gates to perform AI compute, bringing the compute power of servers onto a single chip.
I had the fortune of meeting Mythic CEO, Mike Henry and his cofounder, David Fick, through an introduction by an eager potential customer. She was looking for Siri-like speech recognition and natural language processing in the cars she was making — without having to run a cluster of high-power GPUs, which would compromise cargo room, diminish fuel efficiency, cut back on EV range and require the car to connect to the cloud.
Mike’s team accomplished the impossible by fundamentally changing how you do convolutions in silicon — the basic mathematical operation underlying the execution, and training of neural networks. They proved out their technology by building a GPS receiver that consumes almost 75 times less power than the state of the art.
Mythic has built a world-class hardware and AI software team, and produced demo chips demonstrating the orders-of-magnitudes improvement in action. The company has brought on major international companies as customers and engaged with many more towards entering mass production in the next year.
Moving forward, Mythic needs to show off different sizes, shapes, and colors of its core technology, demonstrate different use cases, and instill confidence in an otherwise conservative community of customers.
If Mythic achieves its goals, it can help bring to life new products that would have been impossible to build without it. Over the next few years, I’m looking forward to seeing what inventors and makers of machines will do with it.