May Use External Foundries For Future CPUs
“Building semiconductors is like playing Russian roulette. You put a gun to your head, pull the trigger, and find out four years later if you blew your brains out.”
— Former Digital CEO Robert Palmer
During its Q2 conference call last night, Intel dropped several bombshells in quick succession. First, the company’s 7nm process node — the same node the company was projecting such confidence about in March — is apparently a full 12 months behind where Intel intended it to be at this point in its development cycle. As a result, the company’s 7nm launches will be delayed at least six months.
The company did not describe the defect in detail, only noting that it had performed a root cause analysis and believes there are no “fundamental roadblocks” to fixing the problem and commercializing the node. In discussing the problem, Swan noted that Intel had developed “contingency plans” to deal with this issue.
If 7nm is delayed, what’s Intel bringing up to replace it? The company had a few words to share on that topic as well. Tiger Lake (mobile) and Ice Lake (Xeon) will both debut this year, as previously announced. Next year, we’ll see Alder Lake, Intel’s first 10nm desktop platform, and a new 10nm server CPU, Sapphire Rapids. Rocket Lake — the supposed 10nm backport built on 14nm and expected late this year — wasn’t mentioned. There’s plenty of evidence that Rocket Lake exists, but it may not have been where Intel wanted to focus attention.
Intel Commits to Using Third-Party Foundries
Ever since Broadwell and 14nm, investors have periodically asked if Intel would tap third-party foundries or consider going fabless. Intel’s response to these questions has generally been that it does tap third-party foundries for some products (which it does), and to state that its nearly unique status as a semiconductor IDM (Integrated Device Manufacturer) gave it an advantage other foundries can’t match. If you want to read the argument, I actually wrote about it back in 2014. Looking back at those articles, I wince. It’s not that I think they were inaccurate when written — it’s that the system they describe went from ticking like a well-tuned clock to knocking like an engine about to throw a rod.
This time, Intel’s entire approach to the topic was different. Instead of reassuring investors that the products built at third-party foundries would be low-cost hardware, Intel openly acknowledged that it would use whatever technology stack was required to deliver performance leadership, be that fully internal manufacturing, fully external, or a combination of the two. CEO Bob Swan emphasized that this plan is part of Intel’s commitment to flexibility and argued that its willingness to develop what he called contingency plans is a sign that the company is determined to deliver maximum value to both investors and customers.
I don’t disagree. At this point, given the challenges Intel has faced with its own manufacturing, the company would be foolish and potentially negligent if it failed to explore every option. That doesn’t change the fact that six years ago, Intel declared its process node leadership would continue on 14nm and into the future, while in 2020, the company CEO spoke of protecting the company’s roadmaps and products from its own process node problems. “We have learned from the challenges in our 10-nanometer transition,” Bob Swan said, “And have a milestone-driven approach to ensure our product competitiveness is not impacted by our process technology roadmap.”
That’s a stark, dramatic shift. Intel talked up a good game last night, with a discussion of “disaggregation” (chiplets) and its plans to take advantage of advanced packaging technology like Foveros in future products. The company isn’t lying when it says that advanced packaging techniques are critical to continuing to advance silicon scaling and improve performance. Ideas like wafer-scale processing are being evaluated again precisely because there’s a need for new packaging techniques to keep the industry moving forward.
CEO Bob Swan declared Intel would be “pretty pragmatic” about deciding where to build parts based entirely on what kind of hardware it needed to bring to market, and did not discuss the difficulty of ramping designs at multiple foundries simultaneously. Intel’s own fabs are known for deploying Intel-specific technology and manufacturing processes intended to maximize the speed of their own parts, not reduce cost for mass customer manufacturing.
Swan is right that this is the most pragmatic thing Intel could do, given its own manufacturing problems, but a lot of investors will be watching the company’s “contingency plans” carefully. It’s fine to tap a third-party foundry for chipsets or Atom. It’s no problem to farm out low-end manufacturing to make more profitable use of existing manufacturing assets. The instant Intel has to activate one of those contingency plans to handle leading-edge “big core” manufacturing because its own fabs can’t handle the job, those massive factories go from a strength to a weakness on the balance sheet.
It’s believable that Intel might bring an improved 10nm process to desktop, because its original 2017 slides always implied that 10nm++ would, in fact, be slightly superior to 14nm for performance and power consumption in desktop power envelops. But again — the company admitted just six months ago that 10nm wouldn’t be the success that its investors have come to expect. Today, Intel sang the praises of 10nm. Three months ago, the company was half-ready to bury it. It is absolutely possible that COVID-19 caused some of this delay, but Intel did not identify coronavirus as a principal reason behind its roadmap slip.
I think what Intel did today was put a very smooth face on a radical corporate realignment. If I might be permitted a bit of poetic license and a Deus Ex: Human Revolution quote: “It’s not the end of the world, but you can see it from here.”
Based on Swan’s own remarks and timelines, Chipzilla has 24-36 months to demonstrate why it should still own its own fabs. By late 2022 / early 2023, TSMC should be shipping 3nm. Even if we assume that Intel’s 7nm is good enough to compare directly to TSMC’s 5nm, that still puts the Taiwanese company a full node ahead.
Is this the end of Intel? Not by half. Intel’s financials are great, the company is tremendously profitable, its data center business continues to grow, and its cash flow is healthy. AMD was in far more trouble after Bulldozer bombed in 2011 than Intel is now, even facing further delays and the question of whether it will remain an IDM over the long term. The company’s process engineers may be struggling, but its CPU design teams are still excellent.
But having a great CPU design team is a necessary but insufficient component of holding the leadership position in CPUs that Intel has long commanded. The company is capable of mounting aggressive comebacks, but if Intel wants investors to see its foundry facilities as a necessary part of the business rather than a drag on its profits, it’s time to pull out all the stops and fix its factories. No, Bob Swan didn’t say that explicitly, today.
He didn’t have to.
- Intel Discontinues Skylake-X HEDT CPUs, With Cascade Lake Barely Available
- Linus Torvalds: ‘I Hope AVX512 Dies a Painful Death’
- x86 Beware: Nvidia May Be Eyeing an ARM Takeover From Soft Bank