Skip to main content

Intel CTO Greg Lavender interview — Why chip maker is spending on both manufacturing and software

Intel is spending billions on manufacturing.
Intel is spending billions on manufacturing.
Image Credit: Intel

Intel has been on a spending spree ever since Pat Gelsinger returned to the company as CEO earlier this year. He pledged to spend $20 billion on U.S. factories and another $95 billion in Europe. Those expenses are scary to investors as they could take a toll on the chip giant’s bottom line, but Gelsinger said he hopes they will pay off over four or five years.

And Intel is making investments in other ways too. In June, Gelsinger brought aboard Greg Lavender, formerly of VMware, as chief technology officer and senior vice president and general manager of the Software and Advanced Technology Group.

I spoke with Lavender in an interview in advance of the online Intel Innovation event happening on October 27-28. In that event, a revival of the Intel Developer Forum that Gelsinger used to lead years ago, Intel will re-engage with developers.

The event will highlight not only what Intel is doing with its manufacturing recovery (after multiple years of delays and costly mistakes). It will also focus on software, such as Intel’s oneAPI technology. Lavender is tasking Intel’s thousands of software engineers to create more  sophisticated software that help brings more value with a systems-focused approach, rather than just a chip-based approach.

GB Event

GamesBeat Summit Call for Speakers

We’re thrilled to open our call for speakers to our flagship event, GamesBeat Summit 2024 hosted in Los Angeles, where we will explore the theme of “Resilience and Adaption”.

Apply to speak here

We talked about a wide variety of subjects across the spectrum of technology. Here’s an edited transcript of our interview.

Above: Greg Lavender is CTO of Intel.

Image Credit: Intel

VentureBeat: Tell me more about yourself. This seems like a very different role for you.

Greg Lavender: I’ve been in the technology industry for a long time, working for hardware companies like Sun and Cisco. In the early days I was a network software engineer for 25 years, writing system software. Always working close to the metal. I have graduate degrees in engineering and computer science. We all get the same courses on Maxwell’s electromagnetic theory and physics. I’m a math geek. But I came up with the growth of the industry, right? Pat is three months older than me. Our careers have kind of tracked along. We’ve both known each other for not quite 14 years.

VentureBeat: What is the task that [Intel CEO] Pat Gelsinger gave you when he brought you aboard?

Lavender: We’ve known each other since I was running Solaris engineering. He was CTO at Intel. Intel launched the Nehalem platforms, if you remember back when that was their first server CPU. We were only shipping AMD Opteron, dual socket, dual core boxes at the time. Pat gave us some money to port it over to the Intel CPU chipsets. We got to know each other and built a trust relationship there. He obviously hired me into VMware and continued that relationship. He knows I’ve got that hardware and software background.

He surprised me when he called me up. I understood the CTO part, but then he also said I’d be the SVP GM of the software group. I said, “How big is that software group?” He said, “Well, we don’t have a software group. We have fragmented parts of software across the company.” In my first 120 days, about how long I’ve been here, I ran a defrag, a disk defrag, and pulled the other 6,000 person software organization together. Everything from firmware to BIOS to compilers to operating systems, all the Linux, Windows, Chrome, Android. All of our system software, all the security software.

I have a big team now. There’s other parts of software going on in the company, but I’m in the driver’s seat for the software strategy and ensuring the software quality for every hardware product we ship.

Above: Intel is focusing on oneAPI to make software creation easier.

Image Credit: Intel

VentureBeat: Is this a smaller percentage of the staff than it would have been in different years? There were things like Intel Architecture Labs and some of the investments that happened in the last decade way outside the chip space. Has that narrowed down again to a smaller percentage of the overall employees?

Lavender: We have a lot, and I’m hiring more. But I’d just say that Pat came in with his eight years at VMware. I was there for half of that. It’s a real software mindset, that the value of software is enabling the open source software ecosystem. Maybe we don’t need to directly monetize our software, right? We can monetize our very diverse platforms.

I’ve spent most of my time here pushing changes into the new compiler system. We just delivered the AMX accelerator code into the Linux kernel, so that when Sapphire Rapids comes out next year we already have the advanced matrix multiplier for machine learning and AI workloads in the Linux kernel. I have a compiler team — I’m sure you’re familiar with the LLVM compiler ecosystem, where all of our new compilers are built on LLVM. We can accelerate our GPUs, CPUs, and FPGAs. It’s a massive set of IP, and it’s IP we give away for free to enable our platforms. We’re contributing to PyTorch, TensorFlow, ONNX. We just updated Intel acceleration into TensorFlow 2.6. That had 8 million downloads just in Q3. We’re enabling the ecosystem for all the developers out there with these accelerated capabilities. We have our crypto library using OpenSSL, accelerated crypto as software.

I think Intel has just failed to tell everyone about all the cool stuff we’re doing. We talk about our chips and our hardware and our customers. We don’t talk about all this great software. We’ve pulled it all together into my org. And I have Intel Labs, 700 researchers at Intel Labs, with all our future software and AI and ML, as well as our quantum computing group. We have this neural computing chip. We just taped out the second version of it. We open-sourced the programming environment for it, called Lava. There were some articles about Loihi 2. That’s our neural processing chip.

VentureBeat: Is some of the investment in software more around the edges of what Intel does? Would that be harder, because there’s so much capital spending going into manufacturing now, with this recommitment to making sure the core manufacturing part of Intel was taken care of? Maybe that leaves less money for software investment.

Lavender: Our view is we need to prime the ecosystem. We need to be open, be trusted. We need to practice responsible AI in all the things we do with our software. My goal is to meet the developers where they are. Historically Intel wanted to capture the developers. I want to enable them and set them free, so that they have choice.

You may be familiar with the SYCL open source programming language, data parallel C++. It’s an extension to C++ for programming GPUs and FPGAs. We have a SYCL compiler built on LLVM. We make that freely available through our oneAPI ecosystem. We have a new website coming online next week, developer.intel.com, where you’ll find all these things. We’ve just been poor about letting the world know about what those investments have already paid for and delivered. Developers would be shocked to know how much of the open source technology they’re currently using has Intel free software in it. It gives them both a better TCO for running their workloads in the clouds, as well as the datacenters or on their laptops.

If anything is lacking, it’s efficient amplification and communication. Just telling everybody, “This is already here.” From my perspective, I just have to leverage it and go further up the stack. We’ve mostly just pushed out software that enables and tickles the hardware. But we’ve been quietly, or relatively quietly, sprinkling all of these accelerated capabilities in all the common open source environments. I mentioned PyTorch. We just don’t talk about it. What I have to change is marketing and communication. We’re going to do this at Intel.

That’s one of the major themes: engaging with the developer community and getting them access to all this cool technology so that they can choose which platforms they want to run on and get that enablement for free. They don’t have to do anything. Maybe set a flag or something. But they don’t have to do any new coding. As you well know, most developers — of 24 million developers, according to some recent data — are up the stack. If you look at the systems people, there’s maybe 1 million. There’s this big group of people in the middleware layer, the dev sec ops people. Maybe not the no-code/low-code developers, the top of the stack. But there are four million enterprise developers just on Red Hat. The fact that I’m pushing stuff into the new compiler ecosystem, pushing stuff into the Linux kernel, into Chrome, means all that technology will be there for all those enterprise developers. I can instantly enable 4 million developers for Sapphire Rapids or Ponte Vecchio GPU.

Intel's Ponte Vecchio is an almagation of graphics cores.

Above: Intel’s Ponte Vecchio is an amalgamation of graphics cores.

Image Credit: Intel

VentureBeat: If you think of things that Intel is getting back to, that maybe it used to do when it communicated through things like the Intel Developer Forum, are there things you expect will be reminders of that?

Lavender: Intel Developer Forum was one of the best tech conferences back when I was at Sun and Cisco. I think it stopped in, what, 2013? Intel Innovation is essentially a relaunch of that theme. “The geek is back,” as Pat would like to say. We were just rehearsing our dialogues for next week. I love it. We’ve grown up together in the industry. I was originally an assembly language programmer on the 8088 and the 8086. Pat and I cut our teeth on Intel as young kids. It’s just so great to be here together at this time given some of Intel’s missteps in the past. We’re in the driver’s seat, and we’re going to steer this massive company into the future.

All those investments we’ve talked about into our fabs and our foundry services business are part of the overall game plan. But if we build all these chips and then don’t have software to make it sing, what good is that? The software is what makes the hardware sing.

VentureBeat: What are some of the messages for people about how Intel has gotten over those missteps in things like the manufacturing process?

Lavender: Pat’s already been out communicating on that and what he’s doing, putting the company’s balance sheet to work to address the world’s lack of capacity to support the demand for semiconductor technologies. When we broke ground in Arizona three weeks ago, there was a lot of press around that. I think you covered Intel Accelerated, where we discussed Ponte Vecchio and how it will use our new process technology, even using TSMC tiles for the Ponte Vecchio general-purpose GPU. We’ve been adopting the new processes we’ve talked about. We’re getting the yields we need. We’re highly optimistic that the industry demand for semiconductor technologies will make IFS a strong business for us. My team, by the way, develops all the pre-silicon simulation software that IFS customers can use to simulate the functionality of their chip before they send it for tape-out.

VentureBeat: I’ve written a few stories from Synopsys and Cadence about how much AI is going into chip design these days. I imagine you’re making use of all that.

Lavender: Being CTO, I get to look across the whole company. That’s one of the advantages of being CTO. I spend a lot of time with the people in our process technology. They’re leading adopters of AI and ML technology in the manufacturing process, both in terms of optimizing yield from each wafer — wafers are expensive and you want to get the most out of every wafer — and then also for diagnostics, for defects.

Every company has silent data errors as a result of their manufacturing processes. As you get to lower and lower nanometer, into angstroms, the physics gets interesting. Physics is a statistical science. You need statistical reasoning, which is what AI and ML are really about, to help us make sure we’re reducing our defects per million, as well as getting the densities we want per wafer. You’re right. That’s the data to physics layer. You have to use machine learning and inference. We have our own models for that, about how to optimize that so we’re more competitive than our competitors.

Above: Intel CEO Pat Gelsinger breaking ground on chip production.

Image Credit: Intel

VentureBeat: If we go back in history some, Nvidia’s investments in Cuda were interesting for breaking the GPU out of its straitjacket, loosening it up for AI. That led to many changes in the industry. Does Intel have its own version of how you’d like to have something like that happen again?

Lavender: There’s at least three parts to that in the way I think about it. Everyone’s interested in roofline performance. Those are the bragging rights in the industry, whether it’s for a CPU or a GPU. We’ve released some preliminary ML performance numbers for Ponte Vecchio. I think it’s on the 23rd of this month that we’ll be submitting additional ML performance numbers for Xeon into the community for validation and publication. I don’t want to pre-announce those, but wait a couple of days.

We’re continually making progress on what we’re doing there. But it’s really about the software stack. You mentioned Cuda. Cuda has become the de facto standard for programming the GPU in the AI and ML space, not just for gaming. But there are alternatives. Some people do OpenCL. Are you familiar with SYCL, the open source effort for data parallel C++? All of our oneAPI compilers compile for CPU, for Xeon and our client CPUs, for GPU and FPGAs, which are also going into network accelerators particularly. If you want to program in C++ with the SYCL extensions, which are up for standardization in the ISO C++ standards bodies, there’s a lot of effort going into writing SYCL as an open source, industry neutral technology. We’re supporting that for our own platforms, and we’d like to see more adoption across the industry.

I’m sure you’re familiar with AMD announcing their HIP, this thing called a heterogeneous programming environment, which is essentially — think of it as a source-to-source translation of Cuda into this HIP syntax for running on their own CPU and GPU. From Intel’s perspective, we want to support the open source community. We want open standards for how to do this. We’re investing, and we’re going to support the SYCL open source community, which is the Khronos Group. We think that provides a more neutral environment. In fact, I’m told you can program SYCL on top of Nvidia GPUs.

That’s sort of step two, once you get competitive at the GPU level. Step three is, what’s the ecosystem that’s already out there? There’s lots of ISVs that are already in these spaces like health care, edge computing, automotive. Everybody wants choice. Nobody wants proprietary lock-in. We’re going to pursue the path of presenting the market and the industry and our customers with choice.

VentureBeat: How open do you want to be? That’s always a good question.

Lavender: We’ll announce this more specifically at Intel Innovation, but the oneAPI ecosystem we’ve talked about — in some sense, the oneAPI name doesn’t mean there’s one single API. It’s really just a brand name. We have more than seven different vertical toolkits for building various things with the technology. We have more than 40 components — toolkits, SDKs, and so on — that make up the oneAPI ecosystem. It’s really an ecosystem of Intel accelerated technologies, all freely available. We’re doing the oneAPI release. We’re accelerating everything from crypto to codecs to GPUs to FPGAs to CPUs — x86 CPUs, obviously, but not necessarily ours. You can use those tools on AMD if you choose.

Our view is to provide the toolkits out there, and we’ll compete at the system level together with our customers, our partners. We’ll enable all the ISVs. It’s not just the open source. We’ll enable the ISVs to use those libraries. It enables anybody doing cloud development. It enables those 4 million enterprise developers on Red Hat. Just enable everybody. We all know about how software eats the world. The more software that’s out there, in the end, cloud to edge — ubiquitous computing, we call it — that enables the advancement of society, the advancement of culture, the advancement of security.

We’re big on pushing our security features in our hardware through those software components. We’re going to get to a more secure world with less supply chain risk from hackers. Even now, machine learning models are being stolen. People spend millions of dollars to train these things, develop these models, and when they deploy them at the edge people are stealing them, because the edge is not secure. We can use all the security features like SGX and TDX in our hardware to create a security as a service capability for software. We can have secure containers. We pushed an open source project called Kata Containers that gets security from our trusted extensions and our hardware through Linux.

The more we can deliver the value of those innovations in our hardware — that most people don’t know about — through the software stack, then that value materializes. If you use Signal messenger for your communications, did you know that Signal’s servers run on Intel hardware with SGX providing a secure enclave for your security credentials, so your communications aren’t hacked or viewed by the cloud vendors? Only Signal has access to the certificates. That’s enabled by us running on Intel hardware in the cloud. The CTO of Signal will be on stage with me as we talk about this, along with the CTO of Red Hat. The CTO of Signal did his undergraduate honors thesis under me on secure anonymous communication over the internet in 2002. I’m really proud of my student and what he’s done.

Greg Lavender came to Intel in June from VMware.

Above: Greg Lavender came to Intel in June from VMware.

Image Credit: Intel

VentureBeat: How do you think about something like RISC-V?

Lavender: It shows that innovation is ever-present and always occurring. RISC-V is another set of technologies that will be adopted particularly, I would think, outside the United States, as in Europe and China and elsewhere in Asia people want alternatives to ARM for their own reasons. It’ll be another open architecture, open ecosystem, but the challenge we have as an industry is we have to develop the software ecosystem for RISC-V. There’s a massive software system that’s evolved over a decade or more for ARM. Either we co-opt that software ecosystem for RISC-V, or a new one emerges. There’s appetite for both, I think. There’s already investment in ARM, but at the same time there’s potential to develop something that’s not tied to the ARM environment.

There are differing opinions. I’ve heard from various people about the opportunity for RISC-V. But clearly it’s happening. I think it’s good. It gives more choice in the industry. Intel will track and see where it goes. I generally believe that it’s a positive trend in the industry.

VentureBeat: As far as what people can expect next week, when it was in person there were so many different kinds of options for deep dives. I guess you may have even more options when you’re doing it online. How would you compare this experience to what people might remember from before about Intel Developer Forum?

Lavender: It’s going to be very interactive, with Pat and myself, Sandra Rivera, Gregory Bryant for the client side, Nick McKeown. Sandra, myself, and Nick are all new in our roles, around 100-plus days. It’s going to be a lively conversation style. I forget the total number, but we have more than 100 “meet the geek” demos. We’ll have some cool stuff, everything from 5G edge robotics to deep learning, AI, ML, obviously graphics. We’re going to show off our new Alder Lake processor. Lots of stuff about various open source toolkits we’ve launched. You may not have heard of IDPK. It’s an open source project we launched. A lot of people are jumping on the bandwagon to offload workloads that traditionally run on the cores to the smart NIC. We have some partners that will be showing up to talk about our technology and how they’re using it.

It’s only a two-day event, but there’s a lot of material packed into those two days. It’s a video format. You can browse around and pick and choose what you want. I think we’re all fatigued of these virtual conferences. We’re trying to make it not just a bunch of talking heads, but more of an interactive dialogue about things we’re doing, about our customers and how they’re taking advantage of it, and then quickly transitioning to live or recorded demos to show that it’s real. It’s not just marketing. It’s real.

VentureBeat: Does this sort of thing make you wish the metaverse was here, that we could make it happen faster?

Lavender: There’s this whole sociological, anthropological conversation to have about the transition we’ve all been through for the last two years. For me, I worked in banking, so I’ve learned to think like a global economist. You can’t help but do that when you’re CTO of a global financial company. I look at these things at more of the macroeconomic level in terms of the likely societal changes. Clearly the shortages in the supply chain and the chokes in the supply chain have shown the insatiable demand for technology generally. Everything we’re doing now is technology-enabled. Can you imagine if we didn’t have Zoom, Teams, whatever? What would that have been like? Obviously this is something in the human experience. We’ve all experienced that.

Above: Intel has 6,000 software engineers.

Image Credit: Intel

But without a doubt, the demand for semiconductors, the demand for software will outstrip the talent, the global talent we have to produce it. We have to get economies of scale. This is where Intel has an advantage. We have those economies of scale more than anyone. We can satisfy more of that demand, even if we have to build factories. We have to accelerate all of that with software. This is why there’s a software-first strategy here. If we’re talking five years from now, it could be a very different story, because the company is putting its mojo back into software, particularly open source software. We’re going to continue to deliver a broad portfolio of technologies to enable that global demand to be met in multiple verticals. We all know software is the liquid. It’s the lubricant that enables that technology to add social and economic value.

VentureBeat: Does it look like 2023 is when the supply chain gets back to its healthier self?

Lavender: I read the same press you read. It seems like it’s a two-year cycle to get there. I’ve read stories about people building their own containers to take over on a ship and collect the parts to bring back. Walking supplies through customs in various countries to get it through the process and the bureaucracy. Right now it seems like a lot of unusual things are happening. I’ve even heard about people receiving SOC packages and they go to test them and there’s actually no guts inside the SOC. That hasn’t happened to us, but these are the stories I’ve read about in the press.

VentureBeat: I would hope that the U.S. government comes around and sees the need to invest in bringing a lot of this back onshore.

Lavender: The CHIPS Act — I’m sure you’re familiar with that. It’s passed the Senate. It hasn’t yet passed the House. I think it’s tied up in the politics of the current spending bill. The Biden administration is trying to put it through. Obviously we’re supporters of that. It’s as good for the industry as it is for Intel. But your guess is as good as mine about geopolitics. It’s not an area that I have any expertise in.

VentureBeat: As far as some futuristic things, I wonder if you’ve thought about some things like Web 3 and the decentralized web, whether that may come to pass or whether it needs certain investments across the industry to happen.

Lavender: There’s a lot of talk. We all think that the datacenter of the future — you may have heard us talk about going from exascale to zettascale. When you get to those scales, to zettascale, it becomes a communications issue. We’ve invested and pioneered in silicon photonics. We can get latencies over distances to a millisecond. That’s quite a distance you can travel at the speed of light.

First off, the innovations in core networking and the edge — it’s not just 5G. I have a new Nighthawk modem from Netgear. I get 400 megabits download. It cost me 800 bucks for that device, but if you’re on a good 5G network, you see the value of it. We’re going to be close to gigabit before too much longer. 6G is going to give you much more antenna bandwidth as well. The bandwidth has to go there before all the other compute density distributes.

I think what you’re talking about is workloads moving not necessarily to the cloud, but away from the cloud and more to the edge. That’s certainly a trend. We see that in our own business and our own growth, in demand for FPGAs and our 5G technologies. Compute becomes ubiquitous. That’s what we’ve said. Network connectivity becomes pervasive. And it’s software-controlled. There has to be software to manage that level of distribution, that level autonomy, that level of disaggregation.

Humans aren’t good about building distributed control planes. Just look at what goes on today. The security architecture that has to overlay all of that — you’ve created a massive surface area for attack vectors. Again, here at Intel we think about these things. We have the capacity and the manufacturing capability to start building prototype technology. I have Intel Labs. That’s 700 researchers. Those are areas we’re discussing as we look at our funding for the next fiscal year, to start exploring these distributed architectures. But most important, back to the software story — I can build the hardware. We can do that. It’s about how you actually manage that at zettascale.

Above: Intel is taking a systems approach to software.

Image Credit: Intel

VentureBeat: You must be happy that Windows 11 has that hardware security feature built in. I think some of these game companies are starting to realize that ring zero access for things like anti-cheat in multiplayer games is important.

Lavender: Windows 11 requires TPM. I have an old Intel NUC that I use for programming. I’ve tried to upgrade to Windows 11 and it told me I needed to buy a new one because I didn’t have the Trusted Platform Module. I asked my colleagues here when the next NUC is coming out. I don’t want to get the currently shipping one. I want one with the new chips. So I’m in line for a beta box.

I just got put onto the Open Source Security Foundation, along with the CTOs of VMware and Red Hat and HPE and Dell. We’re really going to tackle this problem for the industry in that form. From my platform at Intel as the CTO, I want to engage with all my ecosystem partners so that we solve this problem as an industry. It’s too big a problem to solve one-off.

GB Daily - get the latest in your inbox

Thanks for subscribing. Check out more VB newsletters here.

An error occured.