Update (6/1/2020): We’ve republished the story below as a discussion of how Intel’s mobile efforts consumed so much of the company’s attention, yet ultimately came to naught. In the years since this story was published in 2016, Intel attempted to build a competitive 5G modem that would put it on the cutting edge of the new market. Unfortunately, that effort ultimately failed and Intel sold its 5G modem business to Apple. Atom has continued to develop as a low power mobile solution for entry-level Windows systems and Chromebooks.
The observations on foundry models are more salient now than they were in 2016. Back then, Intel’s 10nm was only a little late and the company was still considered to be on the cutting edge of semiconductor tech. TSMC currently occupies that position, though Intel wants to reclaim its crown by 5nm. Back in 2016, we occasionally heard analysts opine that Intel should sell its fabs and embrace the fabless model due to the increased competition from TSMC. Those whispers have only grown louder. I’m not saying I agree with them, but it’s interesting to see how this article frames the TSMC versus Intel fight. Back in 2016, things looked very different than they do today.
The Qualcomm antitrust lawsuit disclosures paint Intel’s failure to gain market share in a very different light, but I have more to say about that in Part 2 of this updated story.
Original Story Below:
Just over four years ago, we reviewed Intel’s first plausible smartphone, the Xolo 900. While it couldn’t match more powerful devices of the day, it was a solid initial effort. More importantly, the Medfield SoC that powered the Xolo 900 was meant to be a beachhead, a harbinger, a sign of things to come.
Four years and billions of dollars later, it’s now clear that the long-awaited ARM-versus-x86 war won’t be fought in the tablet or smartphone market. Intel has canceled all of its publicly announced 14nm smartphone SoCs and most of its 14nm tablet SoCs that would have shipped in Android devices. The company is diving into 5G research and hopes to lead the industry in developing 5G products. But its 3G and 4G modems are still built on the 28nm process node at TSMC, even as competitors like Qualcomm move to 14nm for their own products.
How did Intel manage to spend up to $10 billion and have so little to show for it? There’s no single, simple answer — but there are definitely some smoking guns. If you want to understand what went wrong for Intel, we need to talk about the semiconductor industry as a whole.
Of Fabs and Foundries
The semiconductor industry is dominated by four companies: GlobalFoundries, Intel, Samsung, and TSMC. Three of these companies — GlobalFoundries, Samsung, and TSMC — are known as pure-play foundries or merchant foundries. Their business model is based on manufacturing silicon for other companies, not designing and launching products that they sell themselves. (Samsung does build some custom silicon for itself, but the bulk of its foundry business comes from external customers).
Unlike the merchant foundries, Intel designs and builds both its fabrication plants and its microprocessors. Until the last few years, Intel never took foundry customers. Intel’s recent attempts to break into the merchant foundry business by attracting a handful of high-margin customers weren’t all that successful and have yet to generate significant revenue for the company.
The distinction between merchant foundries and IDMs (integrated device manufacturers) like Intel is a critical part of why Intel’s mobile efforts played out the way they did.
Intel’s manufacturing strategy historically relied on rapidly adopting new process technology. The bulk of the company’s revenue is derived from leading-edge nodes; older facilities were either upgraded or shut down as they became obsolete. The chart below shows Intel’s R&D expenses as a percentage of total sales as compared to Samsung and TSMC:
TSMC pursues a very different optimization strategy. While it also invests in leading-edge semiconductor technology, the bulk of TSMC’s revenue is earned on older technology nodes. The chart below is based on TSMC’s Q1 2015 results:
TSMC’s revenue by process node
As of Q1 2015, 39 percent of TSMC’s revenue was earned on technology nodes it deployed 10-20 years ago. 54 percent of its revenue came from nodes that were in volume production at least eight years ago. Intel made limited use of older facilities to build its chipsets, but its business model is fundamentally different.
For most of 30 years, that difference didn’t really matter. From the 1980s through 2010, Intel beat its low-volume RISC competitors and seized the data center by leveraging the economies of scale it created in the consumer PC market. It crushed its only serious competitor in the x86 market and forced AMD to sell its own fabs in an effort to survive.
Intel drove volume in new markets by exploiting economies of scale it had created elsewhere
The pure-play foundries and Intel worked in parallel tracks, often contending with some of the same problems, but prioritizing and solving them in different ways. Intel built its foundries to rigorous standards using a philosophy it called “Copy Exactly.” It prioritized high yields, focused almost exclusively on microprocessors, and enforced strict design rules. Intel products were designed to be built at Intel foundries with Intel tools and Intel’s established best practices.
In contrast, TSMC and the other merchant foundries designed their process nodes to meet the needs of many different clients. Their fabs prioritize throughput and flexibility while minimizing cost. Both models worked beautifully for their respective businesses — until their businesses started to collide.
The Limits of Technology
Intel’s failure to gain traction in the mobile market highlights the flaws in treating technological progress as a roadmap for corporate success. Despite recent delays and its own decision to abandon its tick-tock model, Intel still owns the most technologically advanced foundries in the world. It’s the only company to have deployed a true 14nm die shrink and it’ll be the first company to deploy a true 10nm node as opposed to a 10/14 or 10/16 hybrid. All of this is true — as is the fact that Intel spent $10 billion (according to Re/code) in mobile with nothing to show for it.
Technology sites, including ExtremeTech, have often discussed x86 versus ARM or AMD versus Nvidia strictly in terms of process node and roadmap. This works extremely well when discussing advances in battery life or performance, but much less well when applied to corporations who suddenly find themselves in direct competition for the first time in decades.
In Part 2 of this two-part series, we discuss how Intel’s business model and development priorities slowed and sabotaged its ability to compete effectively in the mobile world. We also discuss why the myriad ARM-versus-Atom debates failed to capture the real problems plaguing Chipzilla, and why the company saw such limited adoption of its products.
- Intel Core i9-10900K CPU Review: Comet Lake Paints a Target on AMD’s Matisse
- Intel May Have Reserved Its Top-End 28W Ice Lake CPUs Exclusively For Apple
- Intel, TSMC Reportedly in Talks to Build New US Foundries
No comments:
Post a Comment