Global Unichip (GUC) and a host of other Taiwan chip designers are seeing demand for ASICs take off, driven by systems houses that want to differentiate their products for cryptocurrency mining and AI to deliver greater efficiency.
While AI shows long-term potential, GUC’s main ASIC demand so far is for Bitcoin-mining equipment, according to the company. The players in this business are trying to develop ICs rather than using off-the-shelf GPUs, according to the company. Customers are finding that the efficiency of GPUs is not good enough, GUC says.
The Bitcoin-mining business has quickly popped up for foundry Taiwan Semiconductor Manufacturing Co. (TSMC) and other companies in the TSMC ecosystem. GUC customer Bitmain, a privately held Chinese firm that makes Bitcoin-mining hardware and runs its own mining operations, made $3 billion to $4 billion in profits in 2017, according to estimates by Bernstein Research.
New mining equipment vendors have rushed in to capitalize on the boom. The focus now is on developing more customized machines that offer greater efficiency.
“Nowadays, GPUs are used as ASSPs for the AI market,” said GUC Senior Director Lewis Chu, in an interview with EE Times. “But AI is big data, algorithms. If everyone continues to use ASSPs, it means their algorithms are similar. It’s hard for them to do differentiation.”
GUC’s Chu started his career as a chip designer for Motorola and went to TSMC around the time Motorola spun off Freescale. He later joined GUC.
If Amazon or Facebook continues to use off-the-shelf chips, it might not be easy for the internet giants to implement their internal knowhow, according to Chu.
“This is why many data centers would now like to have their own ASICs,” Chu says. “This will be the trend.”
Design houses help limit the risks involved with making a custom chip. The design of an ASIC in the 16nm to 7nm range costs anywhere from $100 million to $400 million today. Most companies staking their brands on a bespoke chip are too risk-averse to take on that cost if the chance of a design failure is high. If the run rate for the chip doesn’t exceed 2 million units, it’s a losing proposition.
Even so, the design business is looking good. Design starts for AI voice-activated device ASICs will increase at a compound annual growth rate (CAGR) approaching 20 percent by 2021, nearly twice the 10.1 percent CAGR of all ASIC design starts between 2016 and 2021, according to a report by Semico Research.
GUC, whose sales soared 31 percent last year, is a spinoff of TSMC and still works exclusively with the world’s biggest foundry. GUC rival Faraday Technology, a spinoff from United Microelectronics Corp. (UMC), says ASICs account for about two-thirds of the company’s business. Other Taiwan chip designers like MediaTek and Alchip Technologies are also chasing the business.
MediaTek this week announced an addition to its ASIC lineup — a 56G serdes IP chip available with silicon-proven 7nm FinFET process technology. MediaTek’s 56G serdes is a high-performance DSP-based solution. According to the company, it demonstrates best-in-class power efficiency, performance and die area. By developing 7nm and 16nm silicon-proven IP, MediaTek said the chip can integrate easily into cutting-edge product designs.
MediaTek’s ASIC services and portfolio cover applications such as enterprise and hyper-scale data centers, ultra-high performance networking switches, routers or compute applications as well as 4G and 5G service provider infrastructure, AI/Deep-learning applications and novel computing applications that require exceptionally high bandwidth long-reach interconnects.
GUC’s first Bitcoin ASIC business came in 2015. By the second half of 2016, the business boomed as the value of Bitcoin soared.
ASIC demand from Bitcoin miners around the world remains strong, according to Global Unichip. The company is cautiously conservative regarding the Bitcoin business because predicting the size of the market depends very much on the price of the cryptocurrency, the company says. A steep plunge in prices could end the Bitcoin-mining business.
“In China, the mining cost for one Bitcoin is around $2,000 to $3,000,” says GUC investor relations project manager Nicole Chan. As long as the value of Bitcoin remains above that cost threshold, the mining business will continue.
A number of GUC’s Bitcoin customers have increased business from last year, according to Chan. The largest of GUC’s customers in the mining business is Bitmain.
“Even though they are the leader in this market, they still want to migrate to more advanced technology,” Chan says.
Global Unichip said it has recently met some Russian companies that aim to make Bitcoin-mining equipment even though the Russian government has banned the business domestically. While China has halted domestic Bitcoin trading, most of the world’s companies making Bitcoin-mining equipment are Chinese.
GUC says it has so much experience with Bitcoin designs that the company has created a standard operating procedure, shortening the lead time. With a netlist, GUC says it can evaluate performance very quickly. The company will sometimes screen out potential customers that lack advanced technology or sound financial information.
“We will evaluate the efficiency of their design,” says Chan. “The efficiency of the Bitmain chip is a good benchmark for us.”
GUC’s Lewis Chu is the key person in charge of screening proposals from potential clients. “Every year, we can do about 30 NTOs (new tape outs),” Chu says. “In the future, 7nm will need more resources.” The demand on internal resources will increase by as much as ten times compared with a 28nm design.”
That’s just one of the reasons why GUC’s headcount is increasing by 20 percent this year compared with a more normal increase of 3-5 percent in past years. The company has had difficulty finding qualified personnel in Taiwan, but after opening an office in Nanjing, China, in August last year, GUC has been hiring more people there. Seed engineers from Taiwan are helping bring the new hires in Nanjing up to speed.
GUC sees AI becoming a new growth driver. The company says most of that business is at the NRE stage and hasn’t yet been recognized as income. GUC divides its AI applications into three segments: training, inference and edge computing.
Customers are primarily from the U.S., Europe and China. Most of the U.S. demand is for training products, while most of the Chinese demand is for inference, according to GUC.
At this point, GPUs still have the lion’s share of the AI business, Chu says, but he expects the ASIC segment to grow quickly in the future.
Investing in an ASIC still isn’t for the faint of heart. AI needs advanced technology, but the NREs at the advanced nodes cost two to three times those at legacy nodes like 28nm, Chu notes.
GUC did its first 7nm design last year and expects to do two or three more this year. Another two to three potential customers at the 7nm node are under engagement, according to GUC.