Today Intel’s CEO Pat Gelsinger has outlined two key changes to Intel policy: one derived from Intel’s plans to offer foundry services to external partners, and the other from Intel starting to outsource its core compute product families in order to get the best product at a given time. Not only is Intel set to offer x86 core IP to customers through its new Intel Foundry Services, but also Intel is looking to creating leadership compute products on external nodes. These are complete 180º turns from how Intel has previously operated.

For the last 20-25 years, Intel has been steadfast in keeping the crown jewels of its product design firmly inside its very protective walls. Over the years, Intel’s x86 designs have mostly led the market in leadership performance and power (except for Pentium 4 and Rocket Lake), and limiting use/production for Intel-only use has enabled the company to improve that design with laser focus, manufacturing not-withstanding. Keeping the cores for internal use only means that neither customers nor competitors were able to see the raw design specifications, and for a long time this has enabled Intel to keep key features, such as its branch predictors, away from all but the most prying eyes.

In a twist to the norm, Intel is now set to dissolve those walls keeping its x86 cores it itself.

First up is Intel’s Foundry Services, a second crack at offering external customers the ability to use Intel’s manufacturing facilities. Idle fabs are costly, and so with IFS, Intel wants to enable a revenue stream while at the time meeting global demand for semiconductors, especially as it pertains to local supply chain security and migrating the world’s semiconductor reliance away from Asia more into the USA and EU. IFS will stand as a separate business unit inside Intel.

As part of IFS, Intel will both offer raw manufacturing services, similar to a standard foundry like TSMC and Samsung, as well as its portfolio of IP to customers. This is a Big Deal™.  Intel will enable a fully vertical model with its IP portfolio, allowing customers to choose from x86 cores, graphics, media, display, AI, interconnect, fabric, packaging, and other critical foundational IP from other sources (such as Arm, RISC-V). The exact way in which customers will be able to license the IP will be announced in due course, but if Intel were to follow the Arm model, then Intel customers will get access to Intel’s 86 core designs.

Arm’s model is bidirectional: core IP and architecture IP. The first allows you to build an SoC with defined cores, while the latter allows you to build your own cores with the instruction set (like Apple does with Arm). When applied to Intel, with the core IP, a customer can build designs based on Intel’s x86 cores with their own or external interconnects, or in different configurations to Intel’s standard model that are more optimized for what that particular customer requires. At the minute Intel is set only to offer core IP licenses, not architecture IP licenses.

If we take this idea and extrapolate, we could very well see x86 cores combined with new memory controllers, active interposers with custom interconnects.

Intel has kind of done this before, although it was very much a walled garden. Intel offered foundry services almost 7 years ago, under then CEO Brian Krzanich, that allowed very select customers to build new SoC designs, with Intel's help, and only for very select pre-approved use cases. In that time, Intel's effort for a proper foundry business was, in Gelsinger's own words, 'weak'. The new model is set to be more open, as far as we're led to believe.

The only question becomes to what extent will Intel offer x86 cores. Will it be the latest cores designed internally, or would they be a couple of generations behind? Will those designs be offered on a variety of process nodes, or just on a singular process node? Would a customer be able to get a core IP license and build it at another fab? This is where the second part of the announcement comes in.

As part of today’s announcement, Intel has stated that it will be expanding its use of third-party foundry capacity. Pat Gelsinger highlighted that it would be leveraging its relationships with TSMC, GlobalFoundries, Samsung, and UMC, to enable the best manufacturing facilities for its leading edge product designs, from communications and connectivity to graphics and chiplets. This builds on the announcements made by former CEO Bob Swan last year in light of Intel's own troubles on its 7nm process. Today's announcements reaffirms Swan's messaging, given that at the time the word 'pragmatic' was used, so while this has probably been in the works in a while, it is good to get a clear confirmation. As part of this announcement, to quote:

‘Gelsinger said he expects Intel’s engagement with third-party foundries to grow and to include manufacturing for a range of modular tiles on advanced process technologies, including products at the core of Intel’s computing offerings for both client and data center segments beginning in 2023’

The key phrase here is ‘core of Intel’s compute offerings’. It could be interpreted in two ways: at the core of a CPU design is a CPU core, which would mean an x86 design unless Intel were to skew away from x86 (unlikely). The other alternative could be an IO chiplet, which is also a ‘core part’ of a compute offering. Paul Alcorn from Tom’s Hardware has confirmed from Intel that the key element here is ‘compute cores’, and although Intel hasn’t specifically said the ISA of those cores, we are set to believe that Intel does indeed mean x86.

This means that other foundries will have access to the floorplans of Intel’s x86 designs, which used to be a big no-no at Intel. Now in saying that, foundries often have strict NDA requirements that stop them sharing designs with customers, as you might expect, but it’s the fact that Intel is even letting another foundry build x86 cores that is the highlight of this announcement.

All-in-all, Pat Gelsinger is enabling a roadmap that allows Intel to pivot, and pivot hard. Steering the Intel behemoth is difficult at the best of times, however Pat’s arrival and enthusiasm has certainly made the company more comfortable in finding where its next generation of revenue is coming from.

 

POST A COMMENT

70 Comments

View All Comments

  • Matthias B V - Wednesday, March 24, 2021 - link

    TigerLake-H will be competitive to Zen3 in perfromance and offers more features. AlderLake will probably be an advantage in every metric to Zen3+ in mobile as the small cores might be a game changer. On Desktop AMD might have an advantage with all big cores but Desktop is a tiny share. It all comes down to servers and SaphireRapids...

    Anyway it doesn't really matter in 2021/22 as market js different and you can sell anything you can get shipped. So all that matters is capacity for the next 12-18 months.

    That's why [besides the time it takes to implement] they focus resources on 2023 and onwards...
    Reply
  • ET - Thursday, March 25, 2021 - link

    Someone's been drinking. Reply
  • JayNor - Friday, March 26, 2021 - link

    Looks to me like Intel will be producing Sapphire Rapids before zen4 chips come out. They are already sampling and the CEO gave q1 2022 production ramp schedule in the most recent presentation.

    The CEO also stated Alder Lake desktop chips will come before the mobile Alder Lake chips ... in 2H 2021. These add DDR5 and PCIE5 according to recent leaks. So, looks like Intel will move ahead in desktop chip technology in mid 2021.
    Reply
  • wut - Tuesday, March 30, 2021 - link

    You're typing like a teen. What a cringe to read... Reply
  • Santoval - Wednesday, March 24, 2021 - link

    This quite frankly sounds like a premature April Fool's joke, particularly the part of core IP licensing for customers to fab x86 parts *not* at Intel. Almost as surprising is that Intel are willing to license x86 IP at all. So they are choosing new revenue streams over control? Wait, you're sure we are talking about Intel right?

    This is even more surprising than announcing that they are going to spin off their fabs ala AMD. This might be the next step, but even AMD did not license their cores, not even the old ones. I strongly doubt they are going to license anything newer than Skylake and its multiple variants; not even Comet Lake, I would guess Coffee Lake tops. Perhaps in 2 - 3 years they will licence Sunny Cove. As for iGPUs they clearly are not going to licence Xe, not even its predecessor Gen11. Crappy Gen9.5 tops. I guess we'll wait and see...
    Reply
  • shadowjk - Wednesday, March 24, 2021 - link

    Does the Chinese Zen count as "licensed AMD cores"? Reply
  • Matthias B V - Wednesday, March 24, 2021 - link

    AMD did not license cores? Well THATIC might be only partially licenced. But what about SONY PS, Microsoft XBox and Samsung future SoC?

    They, especially SONY designs their APUs using AMDs IP with more or less guidance of AMD.
    Reply
  • JayNor - Friday, March 26, 2021 - link

    Looks like Intel wants to get the most out of its 3D fabrication headstart. Perhaps selling x86 tiles for a customer specified system on a package design is their way of fighting off the ARM homebrew projects.

    That doesn't give out their x86 processing secrets ... just a new way to sell cores.
    Reply
  • eastcoast_pete - Wednesday, March 24, 2021 - link

    I read these statements also as a way to address the obvious question about investing billions of dollars now in fabs that won't be ready for years. Having a much more flexible foundry strategy helps assuage the fears that all that capacity will hit just when demand is sagging. By planning for fabrication flexibility already at the chip design and foundry construction stages, Intel is less hemmed in. And, with foundries, the most expensive ones are those that are mothballed. So, fabbing some chips for IBM or fabless RISC-V or even ARM designs helps spread the risk. Being prepared to have TSMC build some CPUs for them means Intel can sell CPUs even if their manufacturing isn't ready. Reply
  • JoeDuarte - Wednesday, March 24, 2021 - link

    This might end up creating opportunities for much more interesting ISAs, like Agner Fog's ForwardCom, the Mill CPU, or Rex Computing.

    RISC-V is pretty boring, an academic project with arbitrary academic decision-making. It's not future-focused enough, and without doing something about the programming language problem, a new conservative RISC ISA doesn't accomplish much.

    There is that crazy performance they allegedly hit last year with a 1 watt core going 5.0 GHz or something, but I haven't seen any follow-up or explanation. An ISA doesn't just give you that kind of performance by itself, especially not a conventional ISA, so I'm not sure what that core was.

    Also, Itanium. Intel should dust it off or something like it. It was good, but they sat around and watched people struggling to build compilers for it. You have to build the compiler with the chip – you can't just introduce the chip and walk away.
    Reply

Log in

Don't have an account? Sign up now