Posted December 17, 2019

This year at CES, Jensen Huang, the CEO of NVIDIA, famously announced that Moore’s Law is dead. We’d argue that it’s not just that Moore’s Law is dead, but it’s also become largely irrelevant.

Let us explain.

Moore’s Law is the tremendously prescient observation made by Gordon Moore, which states that the number of transistors you can fit onto a single chip doubles every 18 months or so. More transistors means more processing power, and this was the trend the industry drafted on for the next four decades, to tackle bigger and bigger problems with ever more powerful applications, until a few significant events that occurred in the last decade. For one, chip manufacturing is reaching the physical limitations of silicon and has caused Moore’s Law to slow. At the same time, the demands of new workloads in the cloud, particularly around machine learning, have required massive increases in processing power. If you look at just the compute requirements for the top AI projects—from AlexNet to AlphaGo—over the last six years, they’ve required an astounding 2x increase in compute every 3.4 months! Far more than the relatively meager 2x increase every 18 months promised by Moore’s Law.

Therefore, even if Jensen is wrong and Moore’s Law isn’t dead, it still isn’t adequate.

In the effort to meet these massive increases in computing demands, the industry has taken the approach of splitting apart the application and running it on multiple chips at once instead of waiting for a single chip to get faster. This is broadly called distributed computing.

But here is the catch. Distributed computing is hard. Really hard. It’s always been a relatively elite subfield of computer science with a much higher bar to achieve than standard programming. Furthermore, distributing an application can greatly complicate a program, making it harder to improve, maintain, and debug.

This is where the open source project Ray comes in. Ray is the leading platform for allowing any developer to write distributed applications, and to do so simply and in a way that is performant, debuggable, and maintainable. Ray allows developers to easily scale up Python applications, and then under the hood, Ray does the heavy lifting around parceling up the work and farming it out to distributed clusters.

Take for example a developer who knows more about machine learning than distributed programming. That developer would build a Python machine learning application as they would a prototype on their laptop. And then with Ray, they would run it across a massive cluster of cloud computing nodes, producing world-class results.

Ray is one of the fastest-growing open source projects we’ve ever tracked, and it’s being used in production at many of the largest and most sophisticated companies on the planet including Ant Financial, Intel, and AWS. While Ray is particularly well-suited for AI and ML workloads, it is used for a wide range of applications across many different industries.

This massive popularity is both a testament to the importance of the problem it is tackling and how well the team behind it has executed on building a product that works and does what it claims.

Ray was developed at UC Berkeley by Robert Nishihara and Philipp Moritz, under the guidance of Ion Stoica and Michael Jordan, and the four of them have co-founded Anyscale, a company to support the development and commercialization of Ray. We’re exceptionally excited to be investors in Anyscale, leading their Series A, and taking a seat on the board.

There is no better team in the world to be tackling this problem. Ion Stoica, a professor at Berkeley, is a world expert in many areas, from core networking and theory, to data processing and distributed systems. He’s a serial entrepreneur, and the cofounder and original CEO of Databricks, another company we’re delighted to be investors in. Robert and Philipp did their PhD with Ion and Michael Jordan. Both are experts in distributed systems and machine learning, and have made major contributions to both the literature in the space as well as architecting, developing, and operating the Ray platform.

Moore’s Law is not going to be the vehicle that keeps the industry apace with the growing compute demands for the next few decades. Traditional approaches to distributed systems, however, just aren’t accessible enough for the general need for more and more compute. We think Ray is the answer and Anyscale is the team to build it. We’re so incredibly excited to be along for the ride.