A Coming Semiconductor Nervana?

By Shahin Farshchi, Ph.D.

Lessons for founders as chip startups emerge from a dark decade, with a vengeance.

For more than a decade, founders and venture investors have steered clear of semiconductors. Their rationale seemed sound. As Moore’s Law brought exponential growth in performance, the cost of designing new chips turned out to far outpace that rate. Bringing a new chip to market can cost well over $100M — therefore, startups, and incumbents alike, need to sell many millions of chips just to break even. Large companies benefit from existing software and product teams to sell these ever-complex products. Startups, however, suffer by having to build up customer support infrastructure for a single product as it slowly ramps. Increasing development costs, compressing margins, slow rates of customer adoption, and industry consolidation have decimated many fledgling companies. Well that’s settled — next industry please!

But investors who did not subscribe to this conventional wisdom (ahem…) were rewarded handsomely with Intel’s acquisition of Nervana, announced earlier today.

I first met Naveen Rao a decade ago in an unsuccessful attempt to recruit him into my Neuroengineering laboratory at UCLA. Naveen had previously spent several years as a chip designer at Sun Microsystems, and later chose to pursue his PhD to learn more about the brain to build brain-inspired chips. He jumped at the opportunity to work with John Donoghue at Brown University, where he built brain-machine interfaces. We were reunited seven years later by my partner Zavain Dar as Naveen set out to raise Nervana’s Series A financing. Along with cofounder Amir Khosrowshahi, Naveen had built a world-class hardware and software team. They aimed to bring powerful software tools to empower developers to bring machine intelligence to the masses — enabled by special hardware (powerful servers to be powered by their unique chips) in the cloud. Nervana’s hardware engineers were among the world’s best digital designers, who invented a means of performing the most common operations used in training deep learning algorithms, natively in silicon.

Nervana’s unique chip architecture accelerated throughput for massively parallelized compute, in silicon. This approach promises an order of magnitude improvement in performance at a given semiconductor technology node. Nervana packaged this technology as a platform for deep learning — Neon — enabling developers to build their deep learning algorithms on the cloud. We invested to empower machines to learn at rates exceeding humans.

Nervana’s story reminded me of the concept of GPUs in the early 1990s; accelerating polygon rendering by doing many multiplications and additions in parallel. Unfortunately, there was a problem: memory was too expensive to realize their full potential. It didn’t take long for memory prices to drop, and the masses were quickly enchanted by the magic of rendering rich polygons in real time, bringing Grand Theft Auto to life. Today, GPUs have outpaced CPUs in PCs, mobile, and gaming, with even more opportunities in new applications in driverless cars, computer vision, natural language processing, and robotic chess players. If Nervana’s chips are what comes after the GPU, why have investors steered clear of semiconductors? To understand how the landscape is different, let’s take a glimpse into the past.

Starting a few decades a go, semiconductor foundries catalyzed the creation of the “fabless” semiconductor startup — no longer did a startup need to raise hundreds of millions of dollars to build a cleanroom facility to build chips. Furthermore, third-party packaging and test companies made semiconductor development a matter of a few clicks in CAD, with fully-packaged chips being drop-shipped to customers. Founders and venture capitalists had a field day with fabless semi home runs such as Qualcomm, Marvell, and Broadcom, just to name a few. The telecom and internet bubble of the ’90s sent their stock prices skyrocketing, leading to a fabless semi startup acquisition feeding frenzy, which prompted more talented founders to start fabless chip companies fueled by venture capitalists eager to generate a quick return.

In the late ’90s and early ’00s, it was a no-brainer for public incumbents to exchange many hundreds of millions in inflated stock for a fabless chip startup, in many cases before they even had a fully-functional test chip. Investors assigned tens of millions of valuation on every PhD doing WiFi, Bluetooth, GPS, SERDES, and fiber-channel. Innovent was sold for hundreds of millions for its Bluetooth know-how. Newport generated a whopping $1B in proceeds after showing a demo chip with less than $5M of invested capital. Unfortunately, the second half of the 2000s flipped the model on its head. The dot-com crash compressed the valuations of the big acquirers. The benefits of Moore’s Law led to skyrocketing development costs, exacerbated by chips getting exponentially more complicated. Competing technologies led to competing standards which took many years to get embraced by big electronics companies — Wireless GigaBit and Wireless USB being examples where most investors in both protocols lost money, with the former eventually being subsumed by Wi-Fi.

Fabless chip companies suffer from a double-whammy: increasing design costs coupled to dwindling gross margins. eSilicon’s 2014 study projects design costs well into the hundreds of millions, while CSI Market Data has observed a constant contraction in gross margin — from over 50% a couple years ago, tumbling down to close to 20% as of mid 2016. Founders need to tilt the playing field in their favor.

Since the mid-2000s, compressed valuations and pursuit of greater economies of scale have brought on industry consolidation — continuing today with Analog Devices’ pending acquisition of Linear. Although the behemoths had greater leverage than startups on customers such as Apple and Samsung, pressure from Asian competitors pushed chip prices — to the extent where Apple captures all the value — both at the retail counter and in the public markets. As a result of the complexities associated with getting larger, the big chipmakers are also getting slower, focusing on iteration rather than fundamental innovation: marginally faster/lower power/cheaper processor and radios. Conventional wisdom accepts chip startups’ fundamental disadvantage, and steers venture dollars away from chip companies altogether.

Large companies targeted startups in the early part of the ’00s. Many of these fabless semiconductor companies were acquired for hundreds of millions after proving out Fiber Channel, GPS, WiFi, Bluetooth, and 3G technology with modest amounts of venture $$, translating to great returns for venture investors. In the latter half of the ’00s, most of large companies (ARM, nVidia and QCOM come to mind as exceptions) saw their valuations get crushed in the public markets while facing stiffer competition from Asia. Depressed valuations, economies of scale, and bargaining power led to large acquisitions over the past decade, making it even more difficult for startups to compete — hence venture running away from chip investments.

Lux’s quest for an unconventional approach to semiconductor startups also led to the creation of Flex Logix. UCLA Professor Dejan Markovic and his Postdoc Dr. Cheng Wang invented a novel interconnect optimization scheme, that when applied to digital circuits, would yield a configurable logic block that outperforms FPGAs at a fraction of the chip area. Dejan and Cheng aspired to build a company around this invention, but feared the capital intensity of building a full-fledged chip startup, and questioned the viability of simply licensing the configurable blocks to chipmakers. Lux introduced Dejan to Geoff Tate, founding CEO of Rambus, which besides ARM, is one of the only semiconductor IP licensing companies that built a market capitalization in the billions of dollars. Geoff recognized three powerful converging themes that would make a Flex Logix incredibly valuable as a licensing company: 1) chip costs soaring, 2) demand for mass customization, and 3) shortening product life cycles. Since partnering together to co-found Flex Logix, Geoff, Cheng, and Dejan have engaged with major semiconductor companies and attracted National Semiconductor cofounder Pierre Lamond — himself having put many semiconductor behemoths in business as a former Partner at Sequoia Capital — as an investor and board member.

Lux continues to seek opportunities in semiconductors, albeit by taking some of the lessons learned over the past decade. Here’s what we look for:

Disruptive technology:

I can feel the eyes rolling. Every VC claims to be seeking disruptive technology, while anyone doing anything in semiconductors views themselves as being on the cutting edge. Attention chip designers: The bar is much higher for you. Founders need to find creative ways of leveraging semiconductor technology currently offered by foundries to offer functionality that is fundamentally unique. It could be a novel circuit topology, or unique algorithm implemented with a unique topology, or do far more with far less. The award doesn’t go to the talented designers that achieve better performance, power, and cost for existing functionality — it goes to those that offer entirely new functionality. For example, basic physics says that the distance data can travel over copper wires is inversely proportional to the data rate — a limitation that light doesn’t apply to light over fiber. As machines got faster, the performance of data centers was limited to power and interconnectivity, resulting from the physical constraints of pushing electrons through copper. Lux-backed Luxtera rose to the challenge of using photonics to connect machines, with the vision of connecting chips, with light. The interconnect bottleneck disappeared, and a new era of high-performance cloud compute will push machine intelligence forward.

Empowering applications:

Incrementally faster and cheaper is great for consumers but doesn’t stand out as an enabler. Chip founders should think less in terms of quantitative performance and more in terms of capability. What does this chip allow you to do that you otherwise can’t do without it? The challenge here is that many applications probably aren’t obvious until designers do something interesting with the chips — in which case I would encourage founders to let their imaginations go wild. Lux helped to lead the spinout of Everspin from private equity-owned Freescale (now NXP) after taking more than $100M of corporate investment to develop its magnetoresistive technology, weaving memory into processing elements. Everspin’s MRAM memory has enabled the world’s fastest solid-state disks, and can empower tomorrow’s in-memory and neuromorphic compute architectures.

Indistinguishable from magic:

Many products we take for granted today would have been nothing short of magical upon their introduction. Broadcom built cheap chips that would push megabits per second over twisted copper wire at a time when conventional technology could barely establish a reliable kilobit per second connection. Qualcomm magically transmitted signals buried in noise with its novel coding techniques. Many engineers have fond memories of the credit-card-like WiFi adapters that magically liberated them from their desks, only later to become standard equipment on all laptops shipped.

At Lux, we’re seeking the brave founders attracting the world’s best talent solving hard problems across all disciplines, including semiconductors, towards building multi-billion-dollar businesses.

Shahin is a partner at Lux Capital. Based in Silicon Valley, he invests in space, robotic, AI, transportation, VR and brain-tech companies; follow him on Twitter @farshchi