Thursday, February 22nd 2024
Intel to Make its Most Advanced Foundry Nodes Available even to AMD, NVIDIA, and Qualcomm
Intel CEO Pat Gelsinger, speaking at the Intel Foundry Services (IFS) Direct Connect event, confirmed to Tom's Hardware that he hopes to turn IFS into the West's premier foundry company, and a direct technological and volume rival to TSMC. He said that there is a clear line of distinction between Intel Products and Intel Foundry, and that later this year, IFS will be more legally distinct from Intel, becoming its own entity. The only way Gelsinger sees IFS being competitive to TSMC, is by making its most advanced semiconductor manufacturing nodes and 3D chip packaging innovations available to foundry customers other than itself (Intel Products), even if it means providing them to companies that directly compete with Intel products, such as AMD and Qualcomm.
Paul Alcorn of Tom's Hardware asked CEO Gelsinger "Intel will now offer its process nodes to some of its competitors, and there may be situations wherein your product teams are competing directly with competitors that are enabled by your crown jewels. How do you plan to navigate those types of situations and maybe soothe ruffled feathers on your product teams?" To this, Gelsinger responded "Well, if you go back to the picture I showed today, Paul, there are Intel products and Intel foundry, There's a clean line between those, and as I said on the last earnings call, we'll have a setup separate legal entity for Intel foundry this year," Gelsinger responded. "We'll start posting separate financials associated with that going forward. And the foundry team's objective is simple: Fill. The. Fabs. Deliver to the broadest set of customers on the planet."
He continues: "We hope that that includes Jensen (NVIDIA), Christiano (Qualcomm), and Sundar (Google), and you heard today it includes Satya (Microsoft), and I even hope that includes Lisa (AMD) going forward. I mean, we want to be the foundry for the world, and if we're going to be the Western foundry at scale, we can't be discriminating about who's participating in that. So, unequivocally, it is to be the foundry for the world. Commit supply chains, your leadership technology - the doors to the à la carte menu are wide open for the industry. So I want my foundry to be used by everybody. Period. We want to help build NVIDIA chips, and AMD chips, and TPU chips for Google, and inference chips for Amazon. Period. We want to help them and give them the most powerful, performant, and efficient technologies for them to build their systems," Gelsinger concluded.
Besides opening up the technological bleeding edge of IFS to third party customers; Intel's medium-term goal is to get into the semi-custom business. Imagine a Foveros-based semi-custom chip that has an Intel Compute tile with its IA cores; a SoC tile designed by Microsoft specific to its game console; and a Graphics tile with NVIDIA GeForce RTX graphics; or a Xeon semi-custom for AWS that has P-core tiles, as well as on-package DPU or network acceleration tiles, and custom security processor tiles. Gelsinger envisions Intel and IFS becoming a single-window for various IP mix-and-match semi-custom chips for customers.
Source:
Tom's Hardware
Paul Alcorn of Tom's Hardware asked CEO Gelsinger "Intel will now offer its process nodes to some of its competitors, and there may be situations wherein your product teams are competing directly with competitors that are enabled by your crown jewels. How do you plan to navigate those types of situations and maybe soothe ruffled feathers on your product teams?" To this, Gelsinger responded "Well, if you go back to the picture I showed today, Paul, there are Intel products and Intel foundry, There's a clean line between those, and as I said on the last earnings call, we'll have a setup separate legal entity for Intel foundry this year," Gelsinger responded. "We'll start posting separate financials associated with that going forward. And the foundry team's objective is simple: Fill. The. Fabs. Deliver to the broadest set of customers on the planet."
Besides opening up the technological bleeding edge of IFS to third party customers; Intel's medium-term goal is to get into the semi-custom business. Imagine a Foveros-based semi-custom chip that has an Intel Compute tile with its IA cores; a SoC tile designed by Microsoft specific to its game console; and a Graphics tile with NVIDIA GeForce RTX graphics; or a Xeon semi-custom for AWS that has P-core tiles, as well as on-package DPU or network acceleration tiles, and custom security processor tiles. Gelsinger envisions Intel and IFS becoming a single-window for various IP mix-and-match semi-custom chips for customers.
23 Comments on Intel to Make its Most Advanced Foundry Nodes Available even to AMD, NVIDIA, and Qualcomm
Doesn't mean they'll use IFS. It's not enough that their process has an advantage. If IFS 18A is 20% better, but you still need to risk an unproven foundry with serious risks of execution issues, treating it's customers badly, and having to move your large and complex multi-billion transistor design along with different details you need to manage, how many will do it? You are risking your entire business potentially crumbling if for any reason IFS fails to execute. TSMC is proven. 20% is nothing. I doubt the advantage will be much greater than that, if any.
This has to be a long game, and geopolitical risks aren't enough. Unless TSMC disappears, even then, because it changes nothing else. If they succeed, it'd be regardless of it, not because of it.
Anyways other than for some old/non critical products why would any of them use Intel Foundry given Intel themselves are rumoured to go with TSMC abondening their own foundry.
Both Intel 4 and 20A are transitory nodes existing solely for the purpose of yield learning for the real one, which is Intel 3 and 18A. Products based on both the transitory nodes will likely underperform, like Meteorlake already is. It's said 20A will be very limited with only product being 6+8 low to mid range Arrowlake-S for desktop and rest being on TSMC N3.
Some are saying Pantherlake will have a Xe3 "Celestial" GPU tile fabbed on Intel 3 process. So when they are ready, they'll move to their own.
It'll take more than transistor leadership to take majority marketshare. Only execution and time will determine it.
I know these all are just speculations. But it's possible, that there' might be some "hidden condition/requirement" in order to get new slice from Act's funding. Why give'em the money, if no other US big company, that needs domestic fab allocation, such as listed above Nvidia, AMD, and Qualcomm benefit from that. Cosnisdering, that each of them is more successful than intel with their products. Seems fair reason.
And the more diverse products they will get, the faster they will refine the foundries. I guess Nvidia could theoretically put all their "consumer" branch on the IFS "rails" as that're not crucial products. But it depends on the price.
The concerns are some technology "leackages" and "migrations" to intel products, to make them more "advance". This might be sensible for AMD, as they have the rivaling CPU branch. But for Nvidia, it's not that bad, as they were more keen to partnership with intel in the past. And Nvidia is now in the situation, when they would benefit from any allocation possible, as their clients don't care, and would buy even literal turds, just be it Nvidia roduct. Even on inferior nodes.
And if intel theoretically become the reliable domestic US supplier, it might put some pressure on both TSMC and Samsung, and finally make some movement in this entrenched way of thnigs, and eventually would benefit every company and consumer. I know it's not TSMC's fault, that other fabs been that incompetent. Taiwan used their tax money wisely, and they got the the incredible results. But intel could also be "TSMC" of US, if the tax money from Chip Act were working properly, and the intel's management didn't make the stupid mistakes. C'mon, that's not that hard.
Just some thoughts.
Assuming they fully open up access to Intel Foundry-specific technologies like Foveros, it could be a way to get rivals on board since TSMC and Samsung basically lets most partners make use of their foundry-specific technologies and iterate on it. Moreover, having other partners design chips that actually iterate on Foveros and other InFo-specific technologies/processes will also help Intel indirectly with their own chips as the processes are refined and approached from other angles Intel might not have been able to investigate alone.
That said, assuming they are serious about playing fair and offering the best to all customers, I could see NVIDIA and AMD potentially using Intel for lower to mid-tier parts, or for specific chiplets where viable, but they probably won't trust them completely with the high-end as long as they're still directly attached to Intel and can't provide a competitive node to TSMC or Samsung.
That said, it would be truly wild to see Ryzen or GeForce fabbed at InFo.
EDIT: Reading the highlights from Toms, it looks like Intel does plan to fully offer their entire set of technologies/processes even to interested rivals, and is also taking steps to ensure legal separation between Intel the Chipmaker and Intel the Foundry so that everyone wins and competes as fairly as possible, while keeping the fabs at maximum capacity too.
Given TSMC's supply constraints and with AMD not being Apple or Nvidia, if Intel's fabs are good enough I would hope AMD makes use of them. We seriously need more competition to TSMC. Samsung has been disappointing to large extent. Hopefully Intel is ultra competitive.
For the second reason, any small performance advantage will be irrelevant. If your product is 20% better but I need to do lot more work, than I won't be using your product, simple.Intel said even their own design team will have to basically compete for Intel foundry service. Gelsinger was talking about how the design team was basically taking advantage of the process teams by doing things a 3rd party foundry wouldn't allow, or would be a waste, for example sending in "hot lots" which is revisions after revisions in a short time rather than waiting and making decisions that reduce such revisions, because each hot lots both waste lot of time and money. They were talking several billions on that alone.
For years people talked about how mediocre CPU design was made up by advanced process technology. The CEO wants the design team to be accountable for their actions now. They have to compete on merit, by designing a competitive product. Basically when the process team failed to execute, it revealed their weaknesses to the public.
The original P4 was ridiculed for high power use and lower than expected performance. The 0.13u Northwood improved that significantly. Let's say if they were stuck on the 0.18u process for little while longer, or the 0.13u didn't provide as big of an advantage?
Also there are products that will be more fit to use on TSMC if the claims by Qualcomm are correct. Some even made a comment about how the dGPU(ARC) was on TSMC because their process was better fit for a GPU. It'll continue to be true until it isn't. Just like they didn't focus on graphics drivers until a dGPU came out, they won't focus on a process fit for low power until a demand is there - foundry is a big driver.
They would probably have to limit theoretical Black Editions to #800X3D only to give it more thermal headroom with only having 1 CCD, unless a theoretical #950X3D Black Edition pairs a high-frequency X3D Zen CCD with a Zen-c CCD for a Hi/Low mix, making use of the high efficiency c cores to provide 8-16 threads for background tasks while using 8 Intel-Fabbed performance cores for mostly gaming.