Intel is welcoming new customers as it reimagines its future as a chip manufacturing powerhouse, but it will have to earn its stripes by proving it won’t repeat past fabrication mistakes that put the company in a tailspin, analysts said.
The chipmaker this week signed up Mediatek as a new customer for its factories, with chips being made on the Intel 16 node (based on Intel’s 22nm FinFET Low Power process). Mediatek is dabbling on this mature Intel node, and if everything works out, could shift over to advanced nodes.
“MediaTek plans to use Intel 16 for the initial engagement on smart edge devices, but we anticipate a long-term partnership that could span multiple technologies and applications,” Jason Gorss, an Intel spokesman, told HPCwire.
Intel’s hope is to get customers like Mediatek on to advanced process nodes and packaging technologies through new factories in the U.S. and Europe. Intel is investing billions in leading-edge factories near Columbus, Ohio, and in Magdeburg, Germany, where high-performance chips could be the earliest products to be manufactured.
Under CEO Pat Gelsinger, Intel has laid out a plan to establish four new nodes in the next five years. That’s a break from the historical trend of advancing a node every 18 to 24 months.
The new nodes include Intel 7 (formerly known as 10nm), followed by Intel 4 (previously called 7nm) in the second half this year, and then Intel 3 in the second half of next year, followed by Intel 20A in the first half of 2024, and Intel 18A in the second half of 2024.
The new factories incorporate new chip design methodologies, transistor technologies and packaging options, which will be available to Intel’s customers. Intel is providing customization and design services to customers.
Intel previously manufactured chips for itself, and Gelsinger’s goal is to open up its factories to fabless companies designing chips. Intel lost its manufacturing leadership to TSMC after a decade of mismanagement and poor manufacturing execution.
Intel struggled with yields on the 14nm and 10nm processes, which stalled the company’s manufacturing advances. Intel management made poor decisions like entering the smartphone market and making 5G modems, and also buckled under competitive pressure from the likes of ARM, AMD, Qualcomm, Nvidia and Apple.
TSMC struck gold luring Apple in as a customer, which brought a higher level of efficiency and discipline in their operations, which ultimately helped the company dethrone Intel as a chip manufacturing leader, said Ruben Roy, senior equity research analyst at WestPark Capital.
“I don’t know if anybody realized how many chips Apple was going to actually develop, design and develop and manufacture. iPhone refresh rates are a lot faster than PCs. PCs can sit out even longer, for maybe three to four years,” Ruben said.
TSMC has since also nabbed Nvidia and AMD as customers, working closely to meet their packaging and chip requirements. Gelsinger is trying to replicate that with its foundry business, in which more engineers will interact with internal and outside entities to react quickly to customer needs and advance faster in the node and manufacturing process.
TSMC hopes to start production of its 3nm chips, which the company calls N3, later this year, and hopes to start making 2nm chips in 2025, companies said in an earnings call this month.
But Intel’s past manufacturing struggles could come back to bite the foundry aspirations. Companies could weigh in on the past before deciding to use Intel’s facilities to make chips.
Gelsinger is talking about catching up and passing TSMC in terms of transistor technology over the next few nodes, and customers could be interested if Intel can offer better or similar technology to TSMC, which would give them an alternate source for chips, said Linley Gwennap, an analyst at TechInsights.
“So far, there hasn’t been a huge amount of interest in that, because Intel’s technology is lagging,” Gwennap said.
Customers may be looking forward to Intel’s 20A and 18A nodes in the 2024-2025 window.
“That’s where people are interested in maybe using Intel as a foundry. But I don’t think people are going to commit to that until they’re pretty confident that Intel can deliver,” Gwennap said.
Gelsinger’s first focus is on transistors, and to make sure Intel can build competitive products on top of those, but won’t be sure if the strategy is working for another year or two.
Customers will also look at Intel’s ability to stick to its roadmap as a measure of execution and reliability of the factories, and while things have been good on the PC side, the datacenter product line is a whole different story.
“Sapphire Rapids is late, the new GPUs are late. And the AI chips from Havana have been behind schedule. It still seems like a big problem for Intel, just to be able to commit to a schedule to deliver things on time,” Gwennap said.
Intel acknowledged the Sapphire Rapids delay on yesterday’s Q2 earnings call. “We have some SKUs out, which is good, but the main SKUs are not out,” said Intel CFO Dave Zinsner. Those higher-volume SKUs will start ramping later this year, but most of the ramp and financial impact will occur next year, Gelsinger added.
Intel has said it needed more time to meet the platform and product validation demands from customers, OEMs and hyperscalers.
The blame game can take any direction, but for customers it’s all about executing well to the roadmap, Gwennap said.
“It seems like there’s problems that crop in at the last minute – they haven’t specced or tested the part properly, or maybe customers want something at the last minute and they’re trying to add it in. [Intel] needs to find a way to deliver these things on schedule, because it just keeps pushing back their competitiveness,” Gwennap said.
Customers who use foundries would typically do validation, testing and packaging, and do test runs and measure yield, said David Kanter,
“TSMC is absolutely the default path and probably the lowest friction. TSMC has done this hundreds of times before,” Kanter said.
“For Intel, working with an external customer is novel, so there could be more bumps along the way,” Kanter said, adding “but anyone worth their salt would anticipate and plan for that.”
But Intel has plenty of experience ramping a process to high-volume, Kanter said.
Mediatek chose Intel 16 for its smart-edge devices because it has some advanced features not commonly found on advanced manufacturing processes, which include a low-cost FinFET node and a slew of new features for RF and analog chips, Kanter said.
Mediatek’s use of an older Intel node could provide it a new chip supplier, but could also be a way to test Intel’s facilities, said James Sanders, research analyst at 451 Research, which is a part of S&P Global Market Intelligence.
“While there’s no mention of specific volumes or revenues in the MediaTek/Intel announcement, the choice of 16nm gives MediaTek a second potential source for chips in – likely – products shipping today. This requires some engineering effort to actualize, though given the realities of the Covid-era supply chain, having options for suppliers provides some breathing room,” Sanders said.
Intel and TSMC have lobbied the U.S. government to pass the CHIPS Act, which passed the House this week (it’s now known as the CHIPS and Science Act of 2022), opening up $54 billion in funding to establish new chip factories in the U.S.
Intel expects to receive some of that government assistance for its new leading-edge factories, but on yesterday’s earnings call Zinsner said it’s too soon to say exactly how it will impact the P&L, which likely would not be until 2023.