5.5.22-Efficient-AI-Sync-Computing

JEFF CHOU: Hello, everyone. Great to be here. My name is Jeff Chou. I'm the co-founder and CEO of Sync Computing and our big mission and motto is let computers provision computers. So what does that really mean?

Today the cloud is fantastic but the process of actually deploying your big data or machine learning job to the cloud is actually still a manual process today. Literally an engineer has to figure out, pick and guess what instances they should use, how should they configure it, et cetera. We think that problem should go away. And there actually is a mathematically best way to use Amazon, or Google, or Azure, and that's our mission. We want to help companies find the mathematically best way to deploy their big data or machine learning jobs possible.

So as everyone here probably knows, the cloud is absolutely huge but there is a massive growing unsustainable problem both in cost and complexity. This report by Andreessen Horowitz shows that it's surpassing $125 billion alone and has surpassed on prem computing itself. So what are we doing at Sync? So at Sync, what we want to do is bridge the gap between low level infrastructure on the cloud and your business goals. So the image on the left shows how the cloud is used today. Its resources versus time. Each one of those gray rectangles you can imagine is one of your big data, or machine learning, or even high-performance simulation jobs that you run on Amazon today.

And like I mentioned, it's an engineer who picks the resources and he or she usually makes over provisions severely. Leading to a lot of waste, and glut, and even poor scheduling. What we want to do is take the exact same resources and jobs that you're running, reallocate resources, even reschedule to give you the best performance, but also align with your business goals. Right? So on the right, you can see, maybe what your business needs is to execute these jobs as fast as humanly possible. You don't care what the cost is. You have to hit your deadlines.

Or maybe these are background jobs, you don't really care if it's 20 minutes or 40 minutes. They just have to get done and you just want to minimize costs, and it's OK if it runs a little bit longer. Or you need some balanced option in the middle. We can basically provide that for you. You just click a button, and it's done. So even if it changes dynamically, you can actually quickly change.

And basically, we let developers focus on code. It's one big sell for a lot of companies, a lot of developers don't really want to work at the low infrastructure level. Two, we give you a whole new level of performance because we're doing, basically, an optimization no human on Earth can do. And three, we let you provision the cloud based on the two things you really care about, which is cost and time. You don't have to understand, what are the different instances? What is the new graviton chip? What is that? It doesn't matter, we'll tell you ahead of time.

So what do we do? How are we different than the rest of the industry? This describes a typical pipeline. I won't get into all the details but basically, on the left you have user code. We don't touch your code. It is what it is. We black box it. We take over these four steps in the middle, such as application tuning, cloud hardware, cloud economics, and then scheduling. There are a lot of companies that focus on isolated elements of this such, as spot pricing.

What we do, we do a global optimization. So how you configure and set your application actually determines what hardware you should pick, but then that's influenced by the spot market, but then maybe you have a crazy schedule and then that machine is unavailable, and you've got to reset all over again. So what we do is, we do all of this at the same time. We basically turn into a giant math problem and just tell you the answer. Then you can enjoy the savings.

In terms of customer results we are running POCs today. The first one was a global streaming company you all probably watched over the weekend. We got them 80% faster and cheaper, running Spark on EMR. Another cloud native company, we got them faster same cost, they're running Databricks today. And Duolingo, an early partner, we got them 55% cheaper. It's interesting, but they also got slower. This is actually one of the more popular options we see customers picking. They have these background jobs. They are not time sensitive, but they run four times or 10 times a day. They just want them cheap, and so we can do that for them.

Right now, a lot of our demonstrations are on single jobs. We wondered well, how does our system scale? So we took Alibaba's cluster trace, and this is 8 days, 14 million tasks, this is their entire cluster. Right? Generates $10 billion in revenue. And we showed that at scale-- This is a simulation, of course. If you apply to our system you could drop their entire usage by 62%, which is obviously a massive, massive number.

In terms of what are we looking for right now, our ideal customer profiles are large companies that run big data workloads, or machine learning workloads in production on the cloud today. That's the only requirement. We don't really care what industry. It could be in finance, SAS, auto, IoT, pharma, research. As long as you're heavy users of the cloud and you care about cost and performance, or accelerating your developers, you could be a customer. Right now we're looking for POC programs, eventually leading to funded pilots.

That's it. I'll be in the cafeteria after this. It'd be great to chat with you all. Thanks so much.