Lip-Bu Tan talks AI, Moore's Law and venture capitalists
SANTA CLARA, Calif. — There’s still room for investing in accelerators for machine learning, one of several semiconductor segments that China hopes to dominate, according to Lip-Bu Tan, a veteran investor. Wearing his hat as chief executive of Cadence Design Systems, he’s bullish on the silicon roadmap and says that EDA companies need to move up to the systems level.
As much as 40% of Cadence’s users are now systems companies of some sort. They are a diverse lot from smartphone makers to top-tier carmakers, mil/aero giants, and the largest data centers that spend an estimated $85 billion a year on capital equipment.
“Quietly, every one of [the big data centers] are designing some form of machine learning or other offload processor … we see more systems companies driving optimizations and seeking differentiation in offering a better experience or more performance or lower latency or more storage capacity,” said Tan in an exclusive interview with EE Times.
Cadence is already working with Medtronic as part of an effort to expand its footprint in medical design. It’s also expanding its work in photonics (Tan spent two days at the recent Optical Fiber Conference in Los Angeles), and it acquired NuSemi, a startup making high-speed SerDes.
“Silicon photonics and high-speed serdes are becoming the bottlenecks in computer networks … NuSemi has a team [that] I have tracked for 10 years,” he said.
The expansion also makes sense given that the number of companies designing leading-edge chips is expected to decline as their costs and complexity rises. That said, Tan is bullish on the silicon roadmap.
“We see tremendous design activity at 7-nm and 5-nm nodes and even have worked with Imec on 3-nm chips … it’s not slowing down; its increasing significantly … I always follow the money, and my three biggest customers are marching to 7 and 5,” he said, noting that 40 customers are working on 7-nm chips.
Despite claims from Qualcomm of diminishing returns at 7 nm, “we see substantial improvements in PPA,” he added, noting that Cadence tools are already optimized for quad-patterning immersion and extreme ultraviolet lithography.
Tan’s take on China’s chip ambitions
The first phase of China’s so-called “big fund” has largely been spent on manufacturing. A second tranche of a similar or larger size is being raised now and will be more focused on “moving up the food chain” to design and intellectual property.
“China’s goal is clear: It wants to be self-sufficient,” said Tan. “In the recently finished central government meeting, semiconductors were number one in the five-year plan, so this goes all the way up to the country’s top leaders.”
China already has a top-five fabless chip company in HiSilicon, the nearly $7 billion semiconductor arm of communications equipment maker Huawei. Strategically, “China wants to be number one in AI; and Alibaba, Baidu, and Tencent are very aggressive in AI development,” he said.
In memory, China’s Yangtze River Storage leads the way with plans to ship 32-layer 3D NAND chips this year. Tan is an investor in GigaDevice, a NOR flash maker.
Overall, the country is still “a couple of generations behind [in memory], but they have huge consumption needs [that] they have to find ways to address,” he said.
China has worked for years to nurture local microprocessor designs. Thanks to deals with the likes of AMD, Intel, IBM, and others, it now has companies producing China-designed ARM, Power, and x86 chips — but none seem to be getting much traction.
“Compute is not everything … I think you will see a sea change as processors become broader, not focused just on computing but on machine learning and other areas … there’s no one path to the processor,” he said, noting that his startup — Annapurna, sold to Amazon — handles offloads of networking tasks.
China’s next big focus on design could spawn its first major EDA company, but Tan does not seem worried. So far, the country only sports two producers of narrow point tools, he said.
China Electronics Corp., a large conglomerate in Guangdong Province that develops one of those EDA tools, is expected to get a $3 billion cut of the big fund soon. For its part, Cadence already has 650 employees in China in cities including Nanjing, where it is setting up an IP subsidiary near TSMC’s planned fab.
“We are open to a joint venture and we are keeping a close eye on China’s EDA plans, but it would take a lot of acquisitions for them to build something similar to a Synopsys or Cadence,” he said.
Hot investments in AI, photonics, materials
With his Walden International hat on, Tan said that VCs have not over-invested in AI accelerators yet. He noted that, at its peak, there were 120 Wi-Fi startups, compared to less than 50 AI chip companies today. Expect more AI startups to emerge, he said.
A CEO of one of his AI investments spoke this week at CDNLive, Cadence’s annual user event. Rodrigo Liang of SambaNova, a former chip and systems executive at Sun Microsystems and Oracle, suggested that the company will make computing appliances to spread AI to business users. If so, it would put the new startup into competition with Wave Computing.
Semiconductor investors have plenty of themes to rally around these days.
“Silicon photonics will be the big bottleneck to go beyond 400G,” he said. “Even the packaging has become a critical issue with insertion loss.”
Looking further out, quantum computing, nanotubes, and new materials are strategic because, as one Cadence executive noted, somewhere around 2-nm traditional silicon will hit atomic limits.
“I’m very interested in gallium nitride for high voltage, RF, analog, and amplifiers … anything to improve PPA and runtime is music to my ears,” said Tan.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times