Page 10 - EE Times Europe Magazine – November 2023

10 EE|Times EUROPE

Intel’s Greg Lavender: ‘We’re Going to Democratize AI’

for autonomous driving. Automotive will be deployment at the edge to accelerate edge natural-language processing.
the biggest demander of silicon, according to inferences to make decisions closer to the For me, there has always been this big
some analysts. application. I think everything from medical, promise that AI was going to take on a lot of
The second-biggest driver is wireless— healthcare, transportation, communications, human activity. Robotic process automation is
mobile Wi-Fi as well as 5G and 6G. The industry and automotive is going to change the most recent version of this notion. It was
wireless buildout continues, and that ever- dramatically in the next five years, not just adopted, but it didn’t fundamentally change
increasing bandwidth allows you do more the next 10 years. the way businesses operate. I know that from
things with your applications. You need But it will be a 10-year run that Intel is when I was CTO at Citigroup.
silicon for the communication chips, and you well-positioned to lead. We have had our
need it to power the more sophisticated appli- challenges, and I’ve worked for the last two Everything is about a systems
cations on mobile devices. years and two months to overcome that—to
The third driver is the cloud and data. Even get our mojo back. We took back our execution architecture—not just selling
before generative AI, there was a grow- discipline. And we now have a software- a chip. That’s a fundamental
ing need to expand cloud computing and defined attitude and strategy in the company.
data stored on the cloud. You now see big We recognize we can’t just sell chips—we have change … and AI is driving it.
demand for bigger and faster cloud comput- to understand how to help customers build
ing architectures that store and process an systems of chips and software platforms at —GREG LAVENDER
ever-increasing volume of data. scale that are often horizontally distributed.
And then the fourth driver is edge com-
puting services. The internet of things and EETE: What other challenges must be We have some impressive results with
automotive are two big applications that will overcome to put AI at the edge? these large language models. But I don’t think
require edge computing services, where the Lavender: AI at the edge is where the action anybody really thinks it’s replacing human
processing is carried out closer to the user, on is going to be with regard to secure AI. People beings. Are the deep cognitive functions of
a device or in a micro data center. are spending a million dollars to train a model humans being mimicked properly? No, they
There will certainly be huge growth in and deploy it at the edge, only to have their are not.
demand, but we think we have what it takes to code and data parameters stolen and then Generative AI is impressive, but it’s differ-
meet that demand. Take for example, the sold on the black market. The edge is a com- ent from artificial general intelligence. I don’t
18A [process] technology we’re driving. By pletely unprotected zone vulnerable to people think we’re there yet.
2030, we’ll have a trillion transistors on a die. wanting to hack AI.
That’s like taking a complete rack of comput- By running in a confidential computing EETE: Given the current state of AI, what
ing power and squeezing it down into a chip. environment, which is Intel’s hardware is the long game for Intel?
We also have the chiplet architecture, Trusted Execution Technology environment, Lavender: The entire industry wants an alter-
which gives you a lot of flexibility. We tend we think we have both the hardware and native to Nvidia. So we’re going to give the
to think of them as little CPUs. But in the software to drive unique capabilities in the market a choice, and we’re going to democra-
future, those chiplets will be much more marketplace. Intel is creating the next inflec- tize AI.
powerful than they are today. If you could tion point, which is confidential computing. We’re following the same strategy we
increase the density of transistors and get This is based on our Sapphire Rapids started when we created the Intel Developer
the right power envelope, chiplets will fun- [4th Gen Xeon Scalable] CPUs, which have Cloud with the goal of putting our latest,
damentally change what you can do on any now exceeded 1 million units sold and greatest chips onto the cloud to give access
device. This is a fundamental driver from the shipped. Another important technology to startups, ISVs [independent software
market perspective. It will enable things like we have is our Trusted Domain Extensions developers] and anybody else who wanted to
edge AI. capability. go kick Sapphire Rapids’ tires before it was
Now, the question is: How do you manage This whole confidential computing concept shipped to our customers. We did the same
all that stuff? The answer is that it will be is still nascent, but it’s growing, and we thing with our GPUs, and now we’re doing it
through software. We can sell lots of chips, expect to monetize all that software and with Gaudi2, the AI chip we got through the
but you need the software to bring it all hardware as people converge confidential Habana acquisition.
together into integrated systems. Every- computing, securing code as it runs in mem- We have a very aggressive roadmap for
thing is about a systems architecture—not ory, and AI. delivering accelerators into the market this
just selling a chip. That’s a fundamental year. Our Gaudi3 processors will soon be
change—one we’re making at Intel. And AI EETE: There is a lot of hype about available for AI acceleration, and then we
is driving it. generative AI. But would you agree that have a new GPU after Ponte Vecchio—other-
What’s driving a lot of this change are there are still some major shortcomings? wise known as GPU Max—code-named Falcon
things like Kubernetes and horizontally Lavender: Absolutely. And that’s why I Shores. We have a growing customer pipeline
distributed computing, which is enabled by think it’s healthy to maintain some level of and growing customer interest.
cloud computing. That’s the foundational skepticism about whether humans are being We will be supporting AI with cloud at the
runtime software. AI workloads cannot run replaced by the latest AI invention. edge using our rich software stack, which
on a single CPU or a single GPU. You have to As a grad student in the mid-’80s, I read was already available for Xeon and is now
gang together a bunch of CPUs and GPUs to a lot of the latest research work on AI. I available for our accelerators. Ours is a
process the data and train AI models in less read all the research that had been done by software/hardware play that will give the
than a year—and sometimes in a matter of Marvin Minsky at MIT; Herb Simon, who won industry a new type of training platform.
several hours, depending on the model size. the Nobel Prize in economics; a lot of the And Xeon is also a very good inferencing
Massive amounts of compute power and work going on at Yale in natural-language platform.
energy are going to get consumed training processing; and Terry Winograd at Stanford. In the end, you have to show up to play.
these models, and then you’ll have this wide Terry Winograd was really the key person in And that’s what we’re doing. ■

NOVEMBER 2023 | www.eetimes.eu