It’s been clear since Pat Gelsinger returned to Intel that he intended to change the company. It’s now becoming clear exactly how. In a far-reaching interview with Pierre Ferragu and Intel’s CFO, George Davis (video embedded below), Gelsinger shared details on what he believes the future holds for Intel, as well as how the company plans to return to overall technology leadership in the next 3-4 years.
One plank of Gelsinger’s turnaround plan for the chip manufacturing giant is known as IDM 2.0. Intel has pledged to launch its own full-featured client manufacturing arm to compete with TSMC and Samsung. Intel has previously announced that it would fab chips for other companies, but the business unit was never a particular focus for the company. Its lack of success, combined with Intel’s manufacturing woes, reflected that fact. This time, the company appears much more serious about the initiative.
“IFS [Intel Foundry Services] gives us a powerful new tool… to go to Amazon and say, ‘Hey, Amazon, you’ve got this Annapurna thing,’ Gelsinger said. ‘Why don’t we fab it for you, why don’t we do some unique packaging, let’s put this together in some unique ways to give you sustainable TCO value that you weren’t able to get before.’”
The Intel executives on the call plainly acknowledged that the company had lost the manufacturing edge it once enjoyed and that Intel no longer commanded the foundry industry as it once had. Intel’s original goal with Atom was to plant x86’s flag in low-power mobile computing, but that did not happen for a variety of factors. Intel’s inability to ramp 10nm from 2015 – mid-2019 was unrelated to its difficulties in the mobile market, but it reinforced the view that the company has lost a step and missed the boat. Gelsinger, however, is more upbeat, especially regarding future AI performance.
“We’re going to start offering more and more products that are going to compete for AI workloads,” the CEO said. “We’ll start knocking Nvidia off the perch they’ve been sitting on for way too long. “We’re going to start delivering competitive products.”
The “competitive products” remark was in regards to both AI inference, which can run anywhere, and AI training, which is typically run on the GPU.
One interesting tidbit that George Davis confirmed: 10nm will be a lower-margin node than 14nm throughout its entire lifespan. “It started in a place and is going to end in a place below the expectations we had for previous nodes,” Davis said, confirming his own previous statements. 10nm yields have improved — Intel should be shipping 10nm hardware in every market segment, including consumer desktop, by the end of the year — but it’s never going to be the earnings powerhouse that 14nm was.
Foundry 2.0 Is the Future
To hear Gelsinger tell it, Intel’s Foundry 2.0 plans are a critical component of the company’s long-term future. It’s all the more striking to hear a longtime Intel vet talk this way because Pat Gelsinger came up at a time when Intel ruled the Earth, as far as semiconductor manufacturing was concerned.
TSMC wasn’t founded until 1987, so the dominant silicon distribution model of the 1980s and 1990s wasn’t for a handful of companies to make chips for everyone else. During this time period, many of the companies that built hardware like HP, IBM, Sun, AMD, and Intel all owned their own fabs. Contract fabbing was not unknown, but it wasn’t common in the workstation, server, and PC businesses. As companies consolidated, the number of competitors shrank. Once AMD sold off its fabs and spun off GlobalFoundries, Intel was the last IDM (Integrated Device Manufacturer) left standing.
Today, Intel has embraced foundry flexibility to a degree that would’ve been unfathomable 25 years ago. During the interview above (the relevant section begins at ~46:53), Gelsinger promised that Intel would “be a big, bad manufacturer of semiconductors for our products, ourselves,” that it would make “selective use” of foundries. The CEO claims Intel can make the decision of which process node to use “late in the design flows, giving us flexibility that the competition doesn’t have.”
If this is true, it seems to represent a shift from what Intel has previously stated. In the past, Intel talked up its legacy as an IDM as incredibly important specifically because it allowed Intel to tune its process node precisely to the requirements of the engineering team, while the engineering team could tune the chip to the characteristics of the process node.
Improving foundry flexibility has been a major goal of Intel’s back at least as far as 2018. Back then, the company announced it would work towards being process neutral in the future, able to deploy products on multiple nodes either at Intel or with other foundries. But it’s not clear if Intel is saying there’s been a tradeoff on the degree to which it customizes hardware or not.
Earlier in the interview, Gelsinger notes that in the past, the fact that Intel hit a node first meant that the actual manufacturing hardware was often tailored for its process and specifications out of the gate, whereas today, companies like TSMC are the ones pushing the industry forward. It may be that Intel performs equivalent customization with TSMC / UMC / GlobalFoundries, or that it can still perform a majority of this work after the design process is completed, regardless of where it builds the final product.
The Future of x86 Licensing Is Still Unclear
The last point Gelsinger reiterates is that future customers will be able to build and customize their own x86 CPUs. He claims there’s a significant market for this capability and that customers wanting to build their own chips are interested in using x86 for backward compatibility.
“So I could go create my own version of Xeon?… I could deprecate some of the transistors I’m not using in those configurations? Yes!” Gelsinger tells Ferragu. But the one thing we don’t know yet is what it means for a customer to “create their own version of Xeon.”
Intel has not stated which x86 IP or core designs it will license or spin off, or whether it will offer an equivalent product to the “hard” core implementations ARM develops for license. ARM doesn’t actually manufacture chips, while Intel does. The “hard” implementation of an Intel CPU core, one could argue, is the actual, physical CPU core and Intel already sells them.
In theory, Intel could be planning to unveil a range of new x86 IP designs to serve these markets, with chip IP licenses and legacy technology requirements that a customer could take to TSMC, GlobalFoundries, Samsung, or UMC. But will it license brand-new instruction sets like AMX (Advanced Matrix eXtensions), AVX-512, or AVX2? Will it license the IP blocks required to build a many-core x86 CPU? If a customer does want a custom Xeon, exactly how custom can the implementation get? Gelsinger confirms that a customer might deprecate some transistors they aren’t using, but does this refer to various I/O blocks and perhaps the ability to optimize the cache, or does it allow for more radical changes? Could a customer license an x86 CPU core based on Skylake, but implemented on Intel 7 with four AVX-512 units?
What happens if an Intel x86 licensee builds a CPU that’s faster or more efficient within a given power envelope than any product Intel currently fields? While this is unlikely to happen in the near future, Intel is not historically known for its willingness to license x86 IP. How Intel answers these questions will partly determine how successful its own foundry efforts are.
What all of this adds up to, however, is scale. Gelsinger wants the scale to move Intel back into overall technology leadership, which means growing Intel’s CapEx, its R&D expenditures, and its customer base. Having recognized that the future doesn’t point towards a world in which every chip is made and branded Intel first and foremost, the company has pivoted to one in which the most powerful, efficient, and useful CPUs and GPUs are either branded “Intel” or manufactured by it.
All of this suggests that IDM 2.0 is much more important to the Intel of today than it was to Intel 6-8 years ago. That’s not enough to win Intel customers in and of itself, but in today’s environment, any foundry with useful capacity is likely to find itself tapped. Intel’s bet is that it can come from behind and overtake the companies that have lapped it in the past half-decade, including both TSMC and AMD. In this context, IDM 2.0 isn’t just one component of a larger strategy. It’s integral to how Intel will win new business and expand the total market for x86. Toss in the fact that Intel is now building its own GPUs and intending to take on Nvidia in that market, and the company’s larger plan for its own resurgence across the semiconductor industry becomes clear.
The strategy Gelsinger has put in play appears diametrically opposed (thus far) to the strategy the company pursued from the late 1980s to the late 2000s. Old Intel guarded its IP jealously and took its largest competitor to the Supreme Court. Old Intel reserved its manufacturing expertise for itself and itself alone. New Intel believes that packaging technologies like EMIB, Foveros, and the upcoming Foveros Direct and Foveros Omnia are key to retaking the position it once held. It appears to be planning to share these technologies with its would-be foundry partners.