Intel IDM 2.0

You've probably read in the press that Intel's new CEO, Pat Gelsinger, laid out his vision for Intel's future last week.

You probably also know that, as he put it, "I've spent four decades in this industry, three of them here at Intel, and now I'm back for my dream job as Intel's CEO". I covered that briefly as part of my post Update: DATE, Achronix, SolarWinds, Batteries, Economist.

So what did Pat say? Well, you can actually watch for yourself (thanks, regulation FD). Here is Intel Unleashed: Engineering the Future. It's one hour long. If you want a recommendation, Ben Thompson of Stratechery calls it "one of the single best keynotes I’ve covered in my career".

Progress Report on 7nm

Another thing that I'm sure you know already is that Intel has struggled with developing the latest nodes of its process (what's called TD, for technology development, inside a semiconductor company, but is a term I generally avoid using since it could mean almost anything outside of a semiconductor context). Its 10nm (roughly equivalent to foundry's 7nm) is finally ramping now in Q1 of 2021. Their next node is 7nm (roughly equivalent to foundry 5nm) has also had problems, although Intel has already said that they are understood and resolved. But 7nm is not due until 2023. Or as Pat put it, "intersecting our 2023 product roadmap".

When Intel initially designed our 7nm process, EUV was a nascent technology so we developed our process to limit the use of EUV, but this also increased the process complexity. As EUV matured and became more reliable, we experienced the domino effects of our 10nm delay, which pushed out 7nm and ultimately put us on the wrong side of the EUV maturity curve. Today I'm happy to announce that we've fully embraced EUV, we've re-architected and simplified our 7nm process flow, increasing our use of EUV by more than 100%. 

Meteor Lake

This is the next-generation server CPU, with a new architecture. The compute tile ("tile" is Intel's name for what everyone else calls a chiplet) will be 7nm. It will also use Foveros (Intel's name for its 3D packaging, see a little more about that in my post HOT CHIPS: Chipletifying Designs where Intel presented Lakefield, which probably gives a hint to how Meteor Lake will come together).

Pat also emphasized something that Cadence (and I in particular) have been saying for some time:

The world will move from system-on-a-chip to system-in-a-package.

Intel actually has two 3D packaging technologies:

• Foveros, which allows die to be stacked vertically (I don't think Foveros stands for anything)
• EMIB, which allows die to be put alongside each other, and stands for Embedded Multi-die Interconnect Bridge

Pat described some details about Ponte Vecchio, which has over 40 different tiles containing over 100B transistors in multiple process technologies (not all of them Intel) using both Foveros and EMIB. Pat even had one to show us, "a petaflop scale AI computer" which is actually destined for Argonne Laboratories exascale supercomputer". Unfortunately, as you may already know, Ponte Vecchio is late by at least six months due to the slip of 7nm, so this Aurora 21 may not be commissioned this year as originally planned.

Manufacturing Strategy

As an IDM, manufacturing is the heart of Intel.  "I call it IDM 2.0," Pat said.

Intel is, and will remain, a major developer of process technology, a major manufacturer of semiconductors, and the leading provider of silicon globally.
...
We will contiunue to build the majority of our products in Intel fabs. We will also expand our use of third-party foundry capacity across our portfolio to deliver the best products in every category we participate in.

Foundry

Today I'm announcing our plans to be a world-class foundry business and a major provider of US and European-based capacity to serve customers globally.
...
We conservatively size the foundry opportunity as a $100B addressable market by 2025

Intel's fabs are in US, Ireland, and Israel. Pat didn't mention it, but they also have a memory fab in China which is in a strange limbo: it was sold to SK Hynix last year, but due to contracts with Micron it is being run by Intel until 2025.

Pat announced the creation of Intel Foundry Services to deliver this vision, a full-vertical, standalone, business unit. Pat also said that Intel will make available a lot of its IP for the foundry business, including x86 cores, graphics, display. AI, interconnect, fabric, and more. Also Arm and RISC-V ecosystem IP.

Although Pat mentioned it and moved on, I think the fact that Intel is making its x86 cores available as IP is a huge deal. When I was at VLSI Technology we had a program to build something akin to a Chrome-book or an iPad, meaning Intel had to let us have the 386 core. We had to keep lists of every individual who had the slightest access to this, and when Intel eventually decided the market was too early and canceled the program, every single storage medium (such as backup tapes) had to be returned to Intel. So I've seen up close just how much those cores are treated like the crown jewels.

Intel has received a positive reaction from a lot of big names. Of course, that's a long way from volume production orders, but knowing how hard it is to get endorsements from major companies like this, there is clearly a demand to diversify at least some manufacturing geographically and reduce the dependence on Asia. Satya Nadella, CEO of Microsoft, showed up in the presentation. Obviously, going back to the Wintel dominance of the PC industry, the two companies have one of the deepest partnerships ever. And, of course, now they are both major players working together in the cloud.

Pat then announced:

Intel's first large-scale foundry operation, which will be in Arizona. We plan to build two new fabs on our Ocotillo campus, with planning and construction starting this year.
...
This represents and investment of approximately $20B.
...
We are excited to be partnering with the State of Arizona and the Biden administration on incentives that spur this type of investment.
...
Today, we are announcing Arizona. And I expect we will be ready to announce the next phase of expansions in the US, Europe, and other global locations within the year.
...
We know our foundry capability must make easy design access and porting from other foundry platforms. To that end, we are making tremendous investments in creating industry-standard PDK models, simplified design rules, and building on our EDA partnerships with Cadence and Synopsys to enable design tools for our Intel Foundry Services.

Pat also committed to deliver the things that Intel knows foundry customers require:

• A separate business unit
• Dedicated capacity
• World-class IP portfolio

Products

Pat then moved on to their product roadmap. I'd almost forgotten that it is an IDM with products, not just a foundry-in-waiting! I'll just call out a few bullet points:

• Tiger Lake (10nm) 30M shipped in over 150 designs
• Alder Lake is sampling
• In two weeks, Intel will announce its third-generation Ice Lake for data centers
• The generation after that for the data center is Sapphire Rapids, sampling now, production end of year, ramping in first half of 2022
• In 2023, Meteor Lake for client and Granite Rapids for data center, with compute tiles on Intel's 7nm

Summary

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.