I had the opportunity to speak with Intel’s CEO Pat Gelsinger before the company’s Innovation 2023 developer conference. It was a chance to get a feel for the advances that Intel has made over the last year and a great way to get a feel for where Intel is going on many fronts, including client computing, data center, edge, and cloud computing. None of these topics can be talked about without the tremendous focus that Intel has put on not only accelerating the AI ecosystem with hardware but also enabling developers to build their AI apps at scale on millions of client devices with Intel’s open-source suite of AI development tools inside of the OpenVINO toolkit.
Executing, Execution, Execution
One of the first things Pat said to me was that he’s glad the company has brought back the spirit of IDF (Intel Developer Forum) after the company “stupidly killed it.” Intel Innovation and Vision are Intel’s ways of reconnecting with the ecosystem in a tangible way that also brings the ecosystem together and helps accelerate ideas and growth. This is part of a broader effort inside the company to improve its execution, which it struggled with prior to his arrival. This is year 2 of a rebuilding process that brings back the systemic engagement that the company once had, and he expects every year of Intel’s developer conferences to be better than the last. Intel gets valuable feedback from these conferences, which ultimately help to drive the company’s future efforts and how it processes industry needs and demands.
Intel’s Pat Gelsinger says that the company continues to execute on time for its “5 nodes in 4 years” foundry strategy and that Intel’s 18A continues to be on track for delivery. This has long been a cornerstone of Intel’s turnaround strategy, which makes sense when you consider Intel’s long-standing strategy of depending on process leadership to deliver leadership performance and power. Intel had a two-plus-year advantage on TSMC; it disappeared. Intel seemed to be a lost company and almost operated like a rudderless ship. Nearly every design needed to be back-ported to an older process. This means product value props and timing change considerably. So, getting back to competitive process nodes and executing on those nodes is critical to the company’s turnaround. It is the most important metric I track. At Deutsche Bank’s 2023 Technology Conference, Pat Gelsinger announced that Intel is not only accelerating its bring-up of its Arizona foundry but has a customer already prepaid for 18A capacity, which is a significant vote of confidence for Intel’s progress. Intel has also partnered with industry leaders like Synopsys to expand its partnership on Intel’s advanced process nodes to enable leading IPs on Intel’s three and 18A process nodes. Considering Synopsys’ leadership in the industry, this is a valuable deepening of a critical relationship for any foundry, especially one like Intel’s, advancing quickly. I am more confident in Intel’s ability to deliver on its promise to execute better with these recent announcements, which I might not have said a year ago.
Some issues are somewhat outside of Intel’s control, like the Tower Semiconductor deal, which got held up by Chinese regulators and has since been abandoned by Intel. Intel has managed, I believe, to side-step that failed deal by signing a US foundry agreement, which includes Tower Semiconductor investing $300 million in its own equipment to help accelerate the bring-up of the IFS fabs that it would end up using. This is a great demonstration of leadership for Intel to turn a situation that it doesn’t have control of and turn it into a net positive for the company and the acquisition target. While the company won’t capture the Tower margin, it’s better than not having the relationship. The key to Intel’s success, as Pat Gelsinger succinctly put it is that the company needs to return to its world-class execution and not only deliver on expectations and deadlines but to exceed them as well.
Given the market explosion of generative AI at the end of last year, many eyes have been only focused on GPU-based generative AI training solutions. The reality is that while that’s where the heat is right now, it likely won’t be the only beneficiary of the recent surge. As we saw with ML and DL, while the initial surge started with training, it led to an inference explosion. Amazon has been quite clear that 80% of its ML budget for Alexa is inference.
I will reiterate from my previous prognostications that Intel is well-positioned in the future of Generative AI. While we haven’t seen much of what I like to call “proof of life, ” evidence is starting to trickle in that Intel could be a competitive and viable alternative to NVIDIA and AMD for AI inference needs. Intel recently submitted three products for this MLPerf round – Gaudi2, 4th Gen Xeon and Xeon CPU Max Series. I have been very critical of vendors for not submitting products, but Intel did.
Results showed that on one workload, Gaudi2 price-performance is significantly superior to NVIDIA A100/H100/GH-200 (near parity on performance alone on GPT-J 6B parameter model against H100). For specific workloads, 4th Gen Xeon CPUs are a decent solution for building and deploying general-purpose AI workloads with the most popular open-source AI frameworks and libraries. What no one wants to talk about is that globally, most AI workloads are run on CPUs. Intel still remains the only server CPU vendor to submit MLPerf results. This is the first time Intel submitted MLPerf results for our Intel Xeon CPU Max series (with up to 64GB of high-bandwidth memory). For GPT-J, it was the only CPU able to achieve 99.9% accuracy
The results show Intel’s competitiveness for AI inference performance on very specific workloads and, I believe, reinforce Intel’s commitment to address the full spectrum of the AI continuum – across hardware and software. But you have to start somewhere.
The company still has a lot of work to do, from winning “specific workloads” to “most workloads,” and I look forward to learning more about the progress at Innovation.
The AI-Accelerated PC
Everyone is talking about AI these days, and that doesn’t exclude Intel. The company is working to enable AI from every front, from the cloud to the PC and everything in-between. Intel is already building AI into every platform, from Meteor Lake on the PC side to Sapphire Rapids on the server/data center side. I expect variants on the datacenter edge like carrier and verticals like retail and manufacturing. Intel has been preparing for years for this moment, and with the introduction of Meteor Lake, the company is preparing to significantly impact the implementation and scale of how businesses and consumers use AI applications. There is still much work to be done at the software level to enable these capabilities, so I believe that companies like Intel, Qualcomm, Apple, and MediaTek must build AI capabilities into their hardware years before it ever gets used.
Developers need to know that the AI capabilities and performance are there before they can even begin to take advantage of the hardware and build AI apps. Generative AI has helped to accelerate some of the developer appetite for AI performance, and we’re starting to see everyone across the industry talk about how they plan to enable on-device generative AI with nimble models that are performant and power efficient. Client-level AI changes the user experience because you can scale in a way that cloud alone cannot. You will still leverage the cloud for model training and retraining, but the industry knows that inference in the cloud is simply too expensive today to scale, especially when talking about applications like generative AI.
In talking to Pat Gelsinger, I got a sense that he believes that the AI PC is a Wi-Fi-like moment where Intel’s Centrino didn’t necessarily define day one of Wi-Fi, but it did enable new use cases and new applications because developers were able to assume that a device would have connectivity. He even references that when Centrino came to market, the 802.11 IEEE standard, which Wi-Fi is based on was already on 802.11g, the fourth iteration of the wireless standard. We are by no means at the beginning or the end of the AI journey for the PC industry, as others like NVIDIA have carried the torch for years with its training, HPC, and cloud focus. But I heard Pat Gelsinger say one extremely controversial claim: “we’re going to deliver more TOPs next year than any other vendor.” He says that they will dwarf Nvidia because they are going to deliver way more volume. I know that with Meteor Lake and Sapphire Rapids will both deliver ample TOPs in volume to validate this claim, but I wonder if companies like Qualcomm who ship in even larger volumes might have some challenges to such claims.
Intel says the company will have hundreds of enabled ISVs on Meteor Lake, which means that I expect to see a lot of developers announcing support for it soon. Intel’s OpenVINO is at the core of Intel’s open-source software capabilities and Intel says that it is gaining broad acceptance across the industry for edge computing. We have yet to see that be the case on the PC, but I believe that Meteor Lake could be the chip and next year might be the time when we really see Intel gain momentum with OpenVINO in the PC. There are a lot of interesting things going on within the industry that I am seeing that lead me to believe that next year will almost certainly be the year of the AI PC.
I do believe that Intel is looking past the current AI bubble and thinking about more than just high-end cloud applications. Intel is the company that will bring us AI at scale across tens of millions of clients, cloud-connected and edge-capable use cases. Intel has always been a market maker company when it is firing on all cylinders and executing. The company was built on its ability to lead and shape the industry around it, which is a foundational aspect of the company’s success and when paired with manufacturing execution and leadership is a healthy recipe for success. Packaging technologies like Foveros are driving some of the interesting new capabilities coming from Intel, like Meteor Lake, and embracing a chiplet architecture is enabling Intel to get back on track as a market maker. Foveros will also be enabling other chipmakers as it has advantages versus TSMC CoWaS and appears to be in higher availability.
I am looking forward to the Intel Innovation 2023 analyst kickoff on Monday and the big announcements on Tuesday. Most of all, I am looking for the 1:1 discussions with the company’s C-Suite.