IFI Insights: Challenging Nvidia - Examining the Patents of an Emerging AI Chip Company

A New Processing Power

In the roughly two years since ChatGPT was unleashed on the public, the hype machine has run nonstop. Adopting AI has felt like an existential issue for business ever since. Companies are racing to advance and exploit the technology in an effort to keep up—or, at the very least, not to be seen as a dinosaur.

The biggest beneficiary of this technology gold rush? Chip company Nvidia, which saw demand surge for its graphics processing units—originally engineered for construction and architecture applications, as well as video editing and 3D depictions—a mighty tool for the power-hungry functions of developing AI. Nvidia’s stock price has seen a meteoric rise in two years (nearly tripling this year alone) as its annual revenue has more than doubled. The company’s upward march has no end in sight. The company’s most recent quarterly earnings performance rocketed past expectations: revenue zoomed 94 percent over the same quarter last year, while profits soared 109 percent. “The age of AI is in full steam, propelling a global shift to Nvidia computing,” said founder and CEO Jensen Huang in a statement. “AI is transforming every industry, company and country.”

That’s for sure. In October, Nvidia surpassed Apple to become the most valuable company in the world (Apple is back on top again). Heavyweight chipmakers, most notably Intel, were caught off guard by the AI-chip run and are scrambling to catch up. It’s not going to be easy considering that Nvidia commands 90 percent of the AI chip market—and still can’t keep up with demand. (For additional analysis on the semiconductor industry, see this IFI CLAIMS report.)

But smaller, more nimble rivals are also throwing down the gauntlet. A Sunnyvale, Calif.-based company named Cerebras is one such contender. Others to keep an eye on: Groq in Mountain View and SambaNova in Atlanta, both of which are working on technology around AI inference, the phase that follows AI training. Founded in 2016, Cerebras makes chips as large as dinner plates. For the purposes of AI, bigger is better because large wafers can fit many more transistors needed to work out enormous AI computational tasks. The company has more than $700 million of backing from investors, including Sam Altman, CEO of OpenAI, the maker of ChatGPT (In fact, OpenAI considered acquiring Cerebras in 2017). In late September, Cerebras filed to go public. Timing is an open question because the Committee on Foreign Investment in the United States (CFIUS) is reviewing one of the investors, according to reports. As Cerebras waits for clearance, it announced an important milestone: a new record in molecular dynamics simulations. “A single Cerebras Wafer Scale Engine achieved over 1.1 million steps per second, which is 748x faster than what is possible on the world’s leading supercomputer ‘Frontier,’” according to a company press release.

In the meantime, it’s worth asking: Can Nvidia reign supreme with possible disruptors coming up from behind? Is a company like Cerebras someplace for investors to allocate capital in this important and growing field? IFI CLAIMS decided to take a look at that question from the perspective of patents. Here is what we found.

Patent David vs. Goliath

Judging by sheer number of patents, Nvidia is way out ahead. The company filed 6,234 applications since 2018 and received 2,355 grants. And the trendline shows a nice upward slope, so the patent pipeline and portfolio looks healthy. Cerebras, in comparison, has filed 128 applications during the time period and earned 50 patent grants. As a young company, its trendline is bumpier, with a 2020 spike in applications followed by a steep decline. Even though it’s important to look at a head-to-head comparison, it’s worth noting that the contrast is not exactly fair. Cerebras was founded less than a decade ago, while Nvidia was born in 1993 and has more than a 20-year head start in patenting activity. One of Nvidia’s oldest patents is this computer graphics system displaying multiple dynamic images (US-4954819-A), filed in 1988. It didn’t originate with the company, but was eventually assigned to it.

To put things in better perspective, Nvidia didn’t start to register its own patents until 1997, four years after the company was formed. In fact, Nvidia didn’t really start to gain traction in patent numbers until 2005. And it’s only over the past four years that Nvidia’s numbers have risen dramatically.

In the Same Class

With patents, the more significant comparison of Cerebras and Nvidia would be in the realms of classifications in which both companies are inventing. Nvidia distinguishes itself from other chipmakers with substantial focus on the G06 class, which covers inventions around simulators and image data generation. The company’s top two technology areas are G06N 3/08, which invents in the area of learning methods and G06N 3/045, which covers encoder and decoder networks. Other fabless chip competitors like Qualcomm and Broadcom, for instance, focus on the H04 classifications more heavily, which encompass electric communication techniques. Like Nvidia, Intel concentrates in the G06 area, but mostly in the electrical digital data processing arena. Another of Intel’s top five areas lies under the H01 umbrella which covers basic electric elements like drying and coating, as well as microstructural devices.

But Cerebras concentrates its inventions in the G06 class, just as Nvidia does. Its chief technology area is G06N 3/063—hardware implementation of neural networks. Cerebras also specializes in G06N 3/084 (learning methods using gradient descent) and G06N 3/045, the same encoder and decoder field that Nvidia occupies. Cerebras might not have the same patent quantities that Nvidia has, but the company is definitely going after the same areas of specialty. For that matter, so are SambaNova, which has been patenting since 2020 and Groq, filing inventions since 2018. Groq’s patents, in fact, are deeply involved in reducing power consumption in the energy-gobbling supercomputing needed to develop AI. And being able to perform faster and better, using less energy, is a compelling competitive advantage. It can take just one great invention to change the playing field, and this is a playing field that has a lot more room for others to grow; so it’s certainly wise for Nvidia not to become complacent—like any great company, it needs to look over its shoulder at the competition.

Issuing a Patent Citation

Imitation is the sincerest form of flattery, and in the world of invention, it’s an adage that holds very true. Creation doesn’t occur out of nowhere. It’s inspired by other innovations and built upon the backs and shoulders of other creators, who can be supplanted. Such is the practice in patenting, which cites the compositions of other patents. This accelerated deep learning patent from Cerebras, granted this year, (US-20200005142-A1) cites this patent from Nvidia (US-20140181501-A1), a heterogeneous multiprocessor design for power-efficient computing, granted in 2017. In fact, this same Nvidia patent is cited by Cerebras for some of their other patents as well, so Cerebras is building upon what Nvidia has already fashioned, another sign that the company is trying to gain traction in Nvidia’s market.

By the same token, the Cerebras patent has been noticed and is gaining forward citations from other important players’ inventions, an indication that Cerebras is onto something good. Among the 90 citations for the accelerated deep learning patent: Microsoft, IBM, Qualcomm, Micron, Bank of America, Amazon, and even the aforementioned smaller rival, SambaNova. In other words, it looks like Cerebras could be sitting on a valuable invention.