“If you put $100 into Intel 25 years ago, your investment would still be worth $100.”

https://x.com/TrungTPhan/status/1819174107646513173

Intel recently suffered the largest decline in its stock price in 50 years, an ignominious prelude to a period that has been and, according to Intel CEO Pat Gelsinger, will continue to be, very difficult for the company. Intel’s market capitalization is now less than it was 25 years ago. Investors have sued, alleging that Intel deliberately concealed the problems that led to this ignominious incident.

What Happened? How could Intel, a company whose meteoric rise decades ago helped coin the term Silicon Valley, have fallen so far from grace? The answer to this question is complex, and I can only scratch the surface in a blog post, but I think the root causes can be traced to 1) Intel does not understand software, 2) Intel owns its fabs, with an honorable mention to 3) Intel does not integrate acquisitions well. An early draft of this post added a numbered point 4) Intel missed the boat on AI, but that follows from points 1) and 3).

1. Intel Does Not Understand Software

Intel is the epitome of a hardware company that does not understand software. For decades, they spent a great deal of R&D money trying and failing to build disruptive hardware that advanced the state of the art, only to have those products failed to achieve market success – even when they collaborated with Microsoft, as they did with Itanium. In the 1990s, Intel won a pitched battle for general-purpose CPU clock cycles by designing the Pentium Pro, a RISC-like implementation to execute their CISC architecture.

The following table summarizes Intel’s forays into computer architecture, organized by decade.

Era	Product	Comments
1980s	i432	Too CISC
1980s	i860	Too RISC; too tight a tie between architecture and implementation
1990s	Itanium	Slow x86 emulation; success was predicated on advancements in the state of the art of compiler technology; not a good fit with processes in Microsoft’s Systems division
2000s	Larrabee	Intel’s offering to compete with CUDA, Cell, etc.
2010s	Xeon Phi	Larrabee without texture units
2010s	dGPU/Gaudi	Belated attempt to re-enter the discrete GPU business. Raja Koduri was recruited from AMD and recently departed.

Within the x86 realm, Intel’s blindness to software did find expression, but their oversights tended to be more correctable. By the 2010s, their x86 business seemed to be humming, with a “tick-tock” tempo alternating between updates to the microarchitecture to improve instructions per clock, often adding new instructions, and a die shrink with bug fixes and higher performance. More recently, Intel abandoned “tick-tock,” apparently without bothering to replace it with a similarly cogent strategy for chip design and evolution. Finally and most recently, Intel badly botched the introduction of AVX512 to the mainstream x86 architecture.

1a. Intel botched AVX512 rollout

Historically, updates to x86 SIMD instruction sets have had small impacts on the die area (and therefore manufacturing costs), because ever since the 80486 (c. 1989) CPUs have been mostly SRAM; and it has been easy to support backward compatibility by ensuring that every newer chip supports all of its predecessors’ features, often in more efficient and higher-performance ways.

It now seems clear that AVX512 was more disruptive than any other SIMD instruction set extension.

I get it, AVX512 is a beast. It not only doubled the SIMD width again over AVX (making the SIMD registers and operations 64 bytes wide, the same size as a cache line), it added predication in the form of new mask registers that have a rich set of instructions to manipulate them. The instruction set extensions were so rich that Intel chose to separate the initial rollout into at least 6 separate feature bits, depending on how you count. But there are ways for hardware designers to pursue area/performance tradeoffs, as AMD has.

Given Intel’s historic performance in adding new x86 features, it’s not a stretch to say that the recent bobbles with AVX512 have contributed to the company’s woes. AVX512 is the first ISA extension I’m aware of that was rescinded in the field through microcode updates. (I am not an operating system architect, but I can only imagine the difficulties that heterogeneous ISA capabilities visit on an OS scheduler. Perhaps shipping devices with some cores that were AVX512-capable and some that were not, is another insight into Intel’s poor software discernment.)

I suspect that with AMD having added AVX512 to all cores, with varying performance characteristics, Intel eventually will be forced to follow in AMD’s footsteps once again, as they did on the 64-bit x86 architecture.

2. Intel Owns Its Fabs

Intel was founded as a foundry or “fab”: a company whose primary business was making semiconductor chips. In Intel’s case, their biggest business was memory chips (DRAMs, or Dynamic Random Access Memories), which happened to be the type of chip that fabs used to test their latest processes. Moore’s Law, the famous observation about the exponential increases in transistor density, is named for Intel co-founder Gordon Moore.

Until the mid-1980s, manufacturing x86 chips was a side business for Intel. When they made a strategic decision to focus on the x86 business, and for many years afterward, it was considered a competitive advantage for Intel to own its foundry. They were at the forefront of innovation through the 1990s and 2000s, relentlessly improving transistor densities in keeping with Moore’s Law (named for Intel co-founder Gordon Moore). By vertically integrating chip design and production, they could closely co-design chips and the processes used to manufacture them, and trade secrets become a reasonable alternative to patents for protection of intellectual property.

Through the 1990s, the x86 business grew so quickly that it could fund the incredibly expensive construction of new fabs that implemented the latest manufacturing processes. For much of that time, Intel had so much manufacturing capacity it was a proximate cause of their monopoly on the x86 chip business: there literally was not enough capacity elsewhere on the planet to service demand.

The challenge with fabs, though, is that because they are expensive to build, the way to maximize the return on that investment is to keep them running long after whatever process they were designed to target has become obsolete. (Retooling fabs to manufacture chips with denser geometries seems to be prohibitively difficult.) With the market growing as fast as it did in the 1990s, the problem of utilizing depreciated factory equipment was less pressing, because the newer factories had more capacity. But as Intel was pursuing this strategy, the industry was continuing to develop an alternative business model where so-called “fabless” semiconductor companies could design chips, then contract with fabs to manufacture those chips. TSMC, the Taiwan Semiconductor Company, is now the most famous of these contract firms; other examples include Samsung and (once upon a time) IBM. The benefit of a company making a pure fab play was obvious: as manufacturing processes advanced, the fabs could simply reduce pricing for outdated processes, and customers who did not need the latest, most expensive process would generate demand.

Most graphics chip companies, including NVIDIA, were fabless. The business and technical relationships between fabs and their biggest customers are sensitive and incredibly important. The fabs must rigorously respect the IP rights of their customers, since often they are making chips for companies that are directly competing in the marketplace.

Until the late 2000s, AMD also owned its fabs; but AMD was never able to use its fabs as a competitive advantage. Since fabs like TSMC had processes to cooperate closely with their customers (and their customers, of course, had experts who could coordinate with fabs on production issues), having the chip designers and the factories making the chips be under the same roof only offered a muted benefit that did not offset the downside risks of having unused, outdated factory capacity.

I actually know an executive who left AMD for NVIDIA in the early 2000s – exactly because AMD still owned its fabs. AMD later was able to divest of its fabs in the Global Foundries deal, which began in 2008 and concluded in 2012.

At some point in the last decade, (I would say more than 5 years ago but perhaps less than 10 years), Intel lost its technology lead over TSMC. They no longer had the best fabrication process in the world, having taken too long to adopt EUV (Extreme Ultraviolet) based lithography methods, famously embodied in ASML’s tin-vaporizing machines. If Intel’s CPU business were fabless, they would be able to access the best fabrication technology to service the business.

Back To The Future

With Gelsinger as CEO, he has made Intel’s future direction clear: he is reversing the course that was set almost 40 years ago, when Intel pivoted from a memory chip company (a fab) to an x86 company. He traveled all over the country, advocating for passage of the CHIPS Act to re-shore semiconductor manufacturing, and has vowed to retake the lead in semiconductor fabrication. Intel is making steady progress, but as their latest quarterly results show, the transition will be a painful one, not least because Intel historically has been much better at making chips for itself than making chips for other companies. Gelsinger is widely expected to lead a divestiture that would echo AMD’s Global Foundries deal, formalizing the division of labor between Intel’s CPU and fab businesses.

3. Intel Does Not Integrate Acquisitions Well

Contrast the acquisition of Altera with AMD’s acquisition of Xilinx: nine years ago, Intel acquired Altera, only to decide to spin it off as an independent entity again recently. In contrast, Xilinx is so fully integrated with AMD that many of CEO Lisa Su’s most trusted lieutenants came to AMD via Xilinx.

Contrast the acquisition of Nervana with AMD’s acquisition of nod.ai. Intel spent $400M acquiring Nervana, then utilized none of its personnel or intellectual property to advance its business or technical objectives. AMD acquired nod.ai, whose former CEO now reports directly to Vamsi Boppana, the ex-Xilinx executive now in charge of AMD’s AI roadmap.

AMD uses acquisitions to increase revenue, broaden the diversity of revenue sources, and deepen its bench of executive management talent. Intel used acquisitions to stage bonfires fueled by their investors’ money.