Intel's Raja Koduri is slated to present a "1000X More Compute for AI by 2025" presentation at Samsung's Advanced Foundry Ecosystem (SAFE) Forum next week, which comes as Intel mulls its strategy for outsourcing some of its production to third-party fabs. The presentation comes after Raja Koduri tweeted an image of a visit to Samsung's Giheung plant in Korea last year, sparking rumors that Intel would use Samsung to produce components for its Xe graphics solutions. Those rumors seemed spurious in the past, but Intel's recent announcement that it will outsource some chip production, and this week's revelation that the company still hasn't decided just what or where it will outsource, makes Koduri's latest interaction with Samsung's foundry all the more interesting.
Intel is certainly at a crossroads. After a decade of dominance fueled by the company's own process tech, the company announced that problems with its 7nm node had forced it to consider outsourcing some components built on leading-edge process tech to third-party foundries, a first for the company. But the company still hasn't developed an outsourcing strategy, as evidenced by Intel CEO Bob Swan's comments in the company's Q3 2020 earnings call. Hence, it's possible we could see the company use either TSMC or Samsung foundries to produce its next-gen flagship chips, or even both.
Swan said that even though Intel will now engage third-party foundries as strategic partners, it will continue to develop its own leading-edge nodes and has deployed a "fix" for its own 7nm node (though that fix has led to an untenable delay). For now, Intel's problem is deciding just where it will build its chips that will come to market in 2023.
Swan said that Intel hasn't decided which chips it will use external foundries for, but did note that "we feel confident in the ability of us being able to port to TSMC," marking the first time the company has mentioned a specific third-party foundry in the context of using it for leading-edge production. That makes it clear that Intel is already working with TSMC, at least in some form.
But there are still plenty of questions to be answered, and that uncertainty is crystal clear in the snippet above from Intel's Q3 earnings deck. Intel says that it is "confident in 2023 product leadership on either Intel 7nm or an external foundry or mix of both."
Intel will likely leverage its packaging tech to reduce the number of externally-sourced components required to build a full chip. Swan says that Intel will decide if it will turn to outside foundries as a stop-gap or invest in its own 7nm equipment, and also where and what to outsource, by "really early next year."
Even though Intel's chips built with externally-sourced components won't come to market until 2023, long lead times require Intel to make a decision soon so it can allow its future partners to build enough production capacity.
But therein lies the problem - any third-party foundry that accepts a contract will likely have to build out significant capacity just for Intel. TSMC is already capacity-constrained and tends to command a premium for leading-edge wafers. Swan also noted that Intel is confident that it can "port back in" from TSMC to Intel's own process tech, meaning the company could move back to its own foundries once it has fixed its own 7nm node. (We imagine those architectures would still require significant re-tuning that wouldn't necessarily classify as a traditional 'port.')
The idea of Intel moving back to its own 7nm probably doesn't seem like an attractive possibility to TSMC, which already has plenty of demand for its 7nm tech. It likely isn't interested in short-term or sporadic business – especially given the large upfront investments required. That means Intel will likely have to commit to procuring a significant amount of capacity from TSMC to secure a contract, much like AMD did with its wafer supply agreement (WSA) with GlobalFoundries.
Meanwhile, Samsung's wafers are relatively less expensive. The company also currently doesn't have as much demand as TSMC, meaning it could have more available production capacity – or at least be willing to dedicate more capacity to Intel's orders.
That leaves plenty of room for Intel to secure at least some of its future chips from Samsung. It certainly doesn't have to be a TSMC "or" Samsung decision: Targeted designs can defray some of the disadvantages of using Samsung's nodes, which are generally less performant than TSMCs, and not all chips have to be on the highest-performing node.
Intel could also seek a licensing deal that allows it to build chips based on an external foundries' process, but in its own facilities. GlobalFoundries employed a similar strategy when it licensed Samsung's 14nm process technology back in 2014, and it's possible the company would be open to a similar deal in the future. Conversely, Intel could also approach TSMC with a similar request, so anything remains possible.
The one thing we do know for sure is that Intel will announce its decision early next year, and it certainly has several options at its disposal. In fact, that's the beauty of embracing the third-party foundry approach: The ability to source different nodes from different vendors based on each product's specific needs affords plenty of flexibility. If Intel fully embraces the third-party model, we could likely see the company source from both TSMC and GlobalFoundries.