切換側邊選單 切換搜尋選單

2017 Computex Taipei AI Forum

Can Taiwan Find an AI Niche?


Can Taiwan Find an AI Niche?


Autonomous vehicles. Robots learning on their own. The artificial intelligence wave has arrived, as was evident when Nvidia founder Jensen Huang spoke at Computex Taipei. Does Taiwan have what it takes to capitalize on this “next big thing?”



Can Taiwan Find an AI Niche?

By Elaine Huang, Liang-rong Chen
From CommonWealth Magazine (vol. 624 )

Not long ago, Sinovation Ventures founder and robotics expert Lee Kai-fu questioned Taiwan’s ability to compete in the artificial intelligence sector, sparking an uproar in academic and industrial circles and fierce debates online.

But he raised some valid questions. Is there any way for Taiwan to capitalize on the approaching AI wave, and what are Taiwan’s strengths and weaknesses in the field? The 2017 Computex Taipei, at which artificial intelligence was one of the featured themes, provided some clues to the answers to those questions.

Some of them came from one of the standard-bearers of the AI revolution, visual computing technology leader Nvidia Corp., which sponsored an AI Forum at the show.

Because Nvidia has rolled out new GPUs (graphics processing unit) capable of supporting the processing speed and power needed for AI operations, Nvidia shares have risen 6.5-fold over the past two years, and the company’s market capitalization has reached US$85.9 billion, not far off from that of mobile phone chipset king Qualcomm Inc.  

That’s why the speech of Nvidia founder and CEO Jensen Huang caused such a commotion before he even uttered a word. The conference room where the AI Forum was being held was completely packed a half hour before Huang was scheduled to speak, and high-tech heavyweights including Advantech Co. Chairman K.C. Liu, Inventec Corp. Chairman Richard Lee and former science minister Shyu Jyuo-min were all in attendance.

“The robot is arguably the ultimate AI,”

“Over the last 30 years, we’ve benefited from what is called Moore’s Law where semiconductor process technology continued to advance and we got 50 percent more transistors every single year. Meanwhile microprocessor architecture continued to advance, taking advantage of those 50 percent more transistors every single year so that the performance continued to increase,” Huang said.  

“Unfortunately, those two dynamics have come to an end,” said Huang, dressed all in black, as he spelled the demise of Moore’s law. Microprocessor performance, he said, “has slowed down to just 10 percent per year.”

 But Huang argued that GPU computing is changing the equation by improving performance exponentially, helping achieve unbelievable “speed-ups” in deep learning and AI applications.

“The growth of deep learning enabled by GPUs has resulted in what is now called ‘the AI revolution,’” he said.

In the AI realm, Huang presented Nvidia’s brand new robotics development platform based on its Jetson TX2 embedded AI supercomputer.

“The robot is arguably the ultimate AI,” he declared.

Beyond making products smaller and more energy-efficient, the platform also enables robot developers to conduct large volumes of simulations and train robots in a virtual reality environment. It eliminates spending on hardware for product testing and works much like AlphaGo, which improved its Go skills by first playing millions of games against itself.

An even bigger attention-grabber was Nvidia’s Taiwan-made advanced data center GPU that runs on chips produced by TSMC’s newest 12 nanometer process and uses TSMC’s Cowos chip stacking technology – the Tesla V100 Accelerator. It not only has the world’s most sophisticated GPU, it introduced “tensor cores” that Nvidia says enable the unit to break speed barriers for deep learning performance.

Perhaps the product most relevant to Taiwan is the open source HGX-I hyperscale GPU accelerator, jointly developed by Nvidia and Microsoft to meet surging demand for AI computing in the cloud. The system has eight Tesla V100 GPUs that can communicate with each other simultaneously and increase its overall computing efficiency.

Huang announced at the Forum that the HGX architecture has been adopted by the four leading server manufacturers in Taiwan – Foxconn, Inventec, Quanta Computer, and Wistron – and they will use the architecture to supply servers to data centers all over the world.

He compared the phenomenon to the last time “Microsoft partnered with somebody to create a standard chassis” when it developed the ATX motherboard in 1995.

“It allowed all of us to focus on one basic chassis and continue to innovate on that chassis, driving its performance up, adding more features, and driving the cost down,” Huang said. “It made it possible for the PC revolution to happen,” just as he is hoping the HGX architecture will lead the AI revolution.

While Intel’s CPU was a bedrock of the PC explosion, this clearly ambitious Taiwan-born entrepreneur is hoping to assume Intel’s mantle as the computing industry undergoes a paradigm shift and set industry standards for data centers with GPUs at their core. He also hoped Taiwanese vendors will continue to play their roles as contract manufacturers.

It would be history repeating itself. When Microsoft and Intel joined forces to create a personal computer platform, it triggered a global personal computer revolution, with Taiwanese manufacturers building the hardware. But because of that division of labor, Taiwanese manufacturers have long been relegated to the bottom of the value chain, gnawing on bones while the two high-tech giants ate meat.

Advantech Chairman Liu argues that the current situation may be different.

“Comparing with the PC contracting of the past, the contracting today is a different game. In the past it was consumer-driven, but today much of it is within vertical supply chains and has higher value and the fields are broader,” Liu said, referring to vertically structured fields such as autonomous cars or smart factories.

As somebody who has turned Advantech into Taiwan’s industrial computer leader, Liu believes Taiwan’s high-tech hardware vendors still offer considerable value in the AI era.

Taiwan Missing from the TPU Pack

Perhaps of concern to Taiwan is that Nvidia has yet to consolidate its perch atop the AI computing field. The new chief technology officer of Intel’s Artificial Intelligence Product Group, Amir Khosrowshahi, called out Nvidia and its reliance on GPU architectures for neural networks that promote machine learning.

“There is so much circuitry in a GPU that is not necessary for machine learning…this is crud that has accumulated over time, and it’s a quite a lot of stuff,” he was quoted as saying by tech website ZDNet.

Among the approaches Intel has come up with is a discrete accelerator called “Lake Crest” that Khosrowshahi told ZDNet is more faithful to the architecture of neural networks.

“It’s a tensor process. It deals with instructions that are matrix operations,” Khosrowshahi said, explaining the concept behind the “tensor processing unit” (TPU), in which the multiplication of matrix elements create operations at a far higher-level than the “element-by-element” multiplication used in GPUs.

To many in the industry, Google has set the bar high with its first TPU that supports Google’s own open source TensorFlow machine learning and artificial intelligence software. By making it open source, Google is hoping to accelerate the advance of the technology.

What’s clear is that some of Taiwan’s venerable high-tech stalwarts have gradually lost the appetite for risk and adventure they once had when faced with a new technology wave, meaning that Taiwan’s AI future may be in the hands of startups.

The IC design field has also jumped into the AI fray. Johnny Shen, the president of Alchip Technologies, which specializes in IC design services, says startups in China, Japan and Israel are working on IC designs for artificial intelligence applications. But because the threshold for entry in the field is high, with development costs capable of reaching US$50 million, “Taiwanese companies have taken a conservative approach and most of them are still in a wait and see mode,” Shen says. 

AI has been recognized within the tech sector as the next big thing after the smartphone, and it presents an excellent opportunity for an upgrade of Taiwan’s high-tech industry. So why are major Taiwanese IC design players like MediaTek Inc. and Sunplus Technology Co. unable to keep pace?

“In fact, Taiwanese vendors were already looking at this a few years ago, but whether or not they have the ability to do this is another question,” says Hung Shih-hao, a professor in National Taiwan University’s Department of Computer Science and Information Engineering. Hung specializes in cloud computing and has been sought out by several Taiwanese companies for help in the field.

Hung says that to create software accelerators for sophisticated machine learning, “the threshold is quite high, much higher than the threshold for mobile phone chips in the past.”

An analyst at a foreign brokerage contends that MediaTek is the only Taiwanese IC design house with the resources to invest in AI chip development, but because mainstream applications such as autonomous vehicles are just getting started, the market scale remains too small to pique its interest.

Taiwanese companies are accustomed to being late movers, the analyst says, so they are naturally reluctant to jump into the field at this stage of development.

What’s clear is that some of Taiwan’s venerable high-tech stalwarts have gradually lost the appetite for risk and adventure they once had when faced with a new technology wave, meaning that Taiwan’s AI future may be in the hands of startups.

At this year’s Computex Taipei, a company created by a group of NTU Information Engineering veterans – Thingnario Ltd. – made a splash with its focus on B2B vertical applications for AI and deep learning. Their platform can use monitors to collect data in a vertical market and then use algorithms to predict the energy consumption of a factory’s machines, motors and generators and when machines need to be maintained.  

“Industrial applications are high-mix, low-volume in nature. Taiwan has an industrial chain, so this B2B model is building on Taiwan’s past strengths,” says NTU Department of Computer Science and Information Engineering professor Winston Hsu, whose specialty is machine learning and AI.

Hsu observes that Taiwan cannot match what companies like Baidu and Google are doing as their point of entry in the AI field. “But it can be like European countries, where although the domestic market is relatively small, they can master technologies and applications in specific industrial fields.”

As Hsu says: “This is the direction Taiwan can go in. There are still a lot of opportunities out there.”

Translated from the Chinese by Luke Sabatier