First things first
Before I start, a quick update: I'm realizing that doing a weekly newsletter involves hitting readers with an awful lot of links at once. Perhaps I'll start doing it twice-weekly.
The burgeoning industry of custom chips for artificial intelligence is the gift that keeps on giving new questions for analysts and, I assume, businesses trying to figure out where to invest resources and energy. I wrote about this issue several times last year (as a reminder, the ARCHITECHT archives are now available here) and, remarkably, the pace of innovation is only picking up. What did we see this week?
- Cloud TPU machine learning accelerators now available in beta (Google)
- Amazon may be developing AI chips for Alexa (TechCrunch)
- MIT’s new chip could bring neural nets to battery-powered gadgets (TechCrunch)
- Soon every gadget is going to have a special chip for AI (Quartz)
As with all things involving AI chips, we need to view this in the context of Nvidia and its surging data center revenues, which I touched on last week. I don't think one can look at everything happening outside of Nvidia (the links above only scratch the surface of this space) and honestly believe that its growth will continue unchecked without the company looking beyond its GPU roots.
Google Cloud TPUs could be the canary in the coal mine here, because they attack the same machine-learning training workloads that GPUs have come to dominate. If users are willing to rent those instead of GPU instances, and if they see significant cost-savings and/or performance improvements, the net result would appear to be that Google will buy fewer GPUs, because demand will be lower. Yes, Google continues to add support for GPUs to products such as Kubernetes Engine as the open source project adds those features, but there's nothing really stopping Google from distinguishing its container products or anything else by also adding TPU support.
Let's say Google is successful with its TPU experiment. Similar results at Microsoft and Amazon -- both of which are at least considering the idea of rolling out their own instances powered by homemade AI chips -- would be a game-changer for AI workloads. In a few short years, we would have went from nothing to GPUs as king, and then from GPU to a range of platform-specific alternatives.
And this is without mentioning whatever effect Intel's Nervana AI server chips will have on the industry. That is something worth thinking about: Even if Nvidia GPUs continue to own training workloads because buyers would prefer hardware platforms that are not locked to specific cloud providers, Intel does have the resources and relationships to make its presence known. Intel also designed its Nervana chips with input from Facebook, which says something about who Intel thinks might be a major buyer of its new hardware.
If I had to make a guess, it would be that Nvidia is willing to take its chances in the data center and just keep turning the dial on GPU innovation. It's already the dominant platform, and dethroning it might be easier said than done.
However, I would also bet that Nvidia makes a few acquisitions on the inference front ("inference" referring to the task of running AI models rather than training them). This where custom Alexa, the MIT research, and lots of other startup and academic efforts come into play. If Nvidia is able to bring in some new technologies and create an integrated platform across both training and inference, it will not only reap the financial rewards of powering billions of AI-powered consumer and industrial devices, but it will also prevent someone else (Intel) from using that tactic to undercut its position.
MongoDB does ACID transactions
This news garnered more attention than I though it would, which I take to be a good sign. It's sometimes easy to forget about MongoDB as we turn our attention on sexy, cloud-native technologies like CockroachDB or time-series databases like InfluxDB (see details on its funding in the Data section below). Or as we watch Oracle try to position itself as a cloud computing leader and promise things like a self-driving database.
However, MongoDB has a large installed base and is performing pretty well as a public company thus far. As companies continue looking for mature databases than can do what they need and also help avoid Oracle or cloud lock-in, a fully featured Mongo could end up looking pretty good in the long run.
CNCF is thinking about serverless computing
The Cloud Native Computing Foundation now has a Serverless Working Group, which as its first order of business published a whitepaper and landscape chart. This makes complete and total sense because, like all computing before it, cloud-native computing operates on a continuum. If containers are on one end today, then serverless is on the other.
But that's almost certainly the direction in which a lot of developers would love to move things. It will be critical for the CNCF and its projects to have a voice in how serverless develops outside of AWS Lambda and other cloud-provider-specific technologies, because there will always be a large contingent of people and companies who prefer to manage these pieces of the stack by themselves rather than outsourcing it. Kubernetes is already the foundation of multiple serverless projects and products, and it seems logical that it, along with other CNCF technologies, could become the de facto foundations for building out and running these internal serverless platforms.
Interestingly, though, one of the more popular open source serverless projects right now is Apache OpenWhisk, which is largely managed by IBM. (Even Google developer advocate and microservices rock star Kelsey Hightower is high on OpenWhisk.) OpenWhisk already supports Kubernetes as a platform for hosting OpenWhisk workloads (yes, serverless does actually require servers somewhere downthe stack), and it will be interesting to watch how the politics of this all play out.