CMOS scaling is not over yet, it’s just slowing down

Source: Content from eettaiwan ， Thanks.

The industry's expected limit of Moore's Law will be the beginning of a transformation in the semiconductor and computer industries. This is the view expressed by industry experts at a panel discussion at a recent event to celebrate the 50th anniversary of the Alan Turing Award.

Although Moore's Law can no longer advance at the same pace, chip, system and software technologies will continue to advance. Experts at the meeting added that without a clear alternative to CMOS scaling, the semiconductor industry and the system industry may form closed islands.

John Hennessy, the 10th president of Stanford University, said: "Moore's Law states that the density of transistors doubles every 18 months. This has been going on for 25 years, but it has gradually slowed down to every two or three years from 2000 to 2005, and has recently evolved to a doubling every four years. Therefore, the industry is gradually approaching the end of semiconductor technology as we expected."

CMOS shrinking “not over, just on hold”?

Another related view, Dennard Scaling, emphasizes that energy requirements will decrease as chips shrink. Hennessy pointed out that Dennard Scaling "has been developing for 10 to 15 years and ushered in the era of dark silicon, which quickly moved to multi-core processors."

In fact, Moore's Law is an observation about economics, not a physical law. The question is whether another physical surface can be found that brings the same return on investment as CMOS, said Margaret Martonosi, a systems expert at Princeton University.

“Moore’s Law is about the rate at which density can be scaled, but we are reaching the end of that at a predictable rate, and we will reach the physical limit in a few generations,” said Doug Burger, an engineer who works on FPGA accelerators at Microsoft’s Azure cloud service.

Norm Jouppi, director of Google's TPU accelerator R&D team, said, "I think CMOS scaling is still a few years away. We will continue to see performance improvements for some related applications in the next decade, but other applications may slow down."

Jouppi sarcastically said that the industry is still in denial about the limits of Moore's Law, just as the shopkeeper in the famous comedy "Dead Parrot" by British comedy group Monty Python said - the parrot "is not dead, it's just resting."

Will DRAM reach its limit first?

Hennessy noted that DRAM may be the first major component to reach its limit. The result "will cause the entire ecosystem to be unbalanced," Burger said.

To overcome the challenge of the memory gap created by DRAM, Jouppi believes that "vertical NAND is the most powerful solution," but investments must be made in many areas.

Flash memory first appeared as a storage application in digital cameras. Today, interface technology has improved significantly, but "there is still a lack of an optimal flash memory computing interface," said Butler Lampson, a distinguished engineer at Microsoft and an adjunct professor at MIT.

Whatever comes next, Burger believes that "the new paradigm in the post-Moore's law era will change dramatically... 20 years from now, the industry will be transformed in an unprecedented way."

Microsoft engineer Doug Burger pioneered graph computing at the University of Texas at Austin (Source: ACM)

Although rising fab costs have led to unprecedented consolidation among chipmakers, "as long as we have three or four stable players, there can be healthy competition," Jouppi stressed, "just like the demand for the iPhone has brought a lot of market pressure."

Both Burger and Jouppi predict an increase in “domain-specific architectures” that deliver the best performance for specific markets. Burger believes this trend will lead to ‘Franken-systems’, “…but the industry is already strong enough to adapt to this trend.”

What if the computer industry returned to vertical integration, with the same company making everything from chips to applications, Hennessy asked, as Apple is heading in that direction and Google appears poised to follow in the same direction?

Still, "there are a number of other architectures that are growing very quickly, some by chip vendors and some by cloud vendors," Jouppi said.

However, will new processor companies emerge after unifying several instruction sets? The panelists had different opinions. Hennessy said: "Chip design requires extremely complex and professional technology, so it is very valuable for processor designers from several companies to work together."

New interfaces and quantum application requirements

The slow end of Moore's Law is also destroying the strategic abstraction layer that software developers have isolated in chips (instruction set architecture). New board interfaces are needed, perhaps for larger vertical markets, but what they should actually be is not yet known.

With an instruction set architecture (ISA), software developers can "make relatively minor changes," but, Martonosi said, "there are about six ISAs in smartphone processors today, and half of the SoC area is accelerators without an ISA."

New architectures and tool libraries are helping fill the gap, but they are creating systems that are “hard to verify and ensure reliability ... which will make the situation worse,” she said, adding that if market-specific systems begin to define their own interfaces, that will require new design flows and engineers who can develop stacked chips.

Burger said cloud computing workloads remain diverse. Cloud providers also have millions of customers, and "even large companies using 1-2% of our server resources can be updated weekly or monthly - but this is still too fast for FPGAs." He emphasized that so-called programmable chips "are still too difficult to program."

Experts at the panel discussion believed that the decline of Moore's Law has created an urgent need in the industry for more efficient software and general-purpose processors.

Jouppi said: "I hope that in the future we can further improve the software and hardware. The industry has not been proactive enough in these two areas in the past. There is still room and enough time to improve efficiency."

“It may not be as good as it was under Moore’s Law, but there is still room for improvement in applications, algorithms and hardware,” said Lampson, one of Alto’s investment firms.

Margaret Martonosi wrote two textbooks on computer power supply effects

Martonosi, who has conducted in-depth research in this field for many years, said that for a long time, there has been no powerful solution to replace Moore's Law, but quantum computers seem to have great development prospects.

“The good news from a physical point of view is that [quantum systems] are very close to reality,” Martonosi said. “You can get a 50- to 100-qubit machine in the next year or two, and someone will be able to program it to give it a speedup over a classical system.”

So far, the bad news has been limited to smaller applications for a few known systems. She added: “There is a huge gap between the number of qubits needed for widespread applications today and the number of systems we can build reliably… and it is still unclear who will buy these thousands of qubit systems in the future and for what applications.”

“We’re going to start building quantum computers, but what big problems they’ll be used to solve — that remains to be seen,” Burger said, noting that he has noticed more programmable biology research based on protein pathways in the future.

Today is the 1319th issue of content shared by "Semiconductor Industry Observer" for you, welcome to follow.

Moore invites you to join the elite WeChat group

Follow the WeChat public account Semiconductor Industry Observation , reply to the keyword in the background to get more content

Reply Ambition , see "Apple's Ambition for a Chip Empire"

Reply Zhang Rujing , read "The Three Falls and Three Rise of Zhang Rujing, the Godfather of China's Semiconductor Industry"

Reply Domestic , see "The biggest beneficiary behind the rise of domestic mobile phones"

Reply to ASR and read "Behind ASR's Acquisition of Marvell MBU: A Love-Hate Story about RDA"

Reply to IC , read "Understanding the IC Industry Structure and Competitive Relationship in One Article"

Reply Exhibition , see "2017 Latest Semiconductor Exhibition and Conference Calendar"

Reply Submit your article and read "How to become a member of "Semiconductor Industry Observer""

Reply Search and you can easily find other articles that interest you!

[About reprinting]: Reprinting is limited to full text reprinting and the title and content of the article must be completely retained. No deletion or addition of content is allowed to circumvent the original protection, and the beginning of the article must indicate: Reprinted from the WeChat public account "Semiconductor Industry Observation icbank". Thank you for your cooperation!

【About Call for Papers】: Semiconductor elites are welcome to contribute (including translation and editing). Once accepted, they will be published with your name, and you will be rewarded with a red envelope! Signing a contract to become a column expert will earn you a thousand yuan in manuscript fees! Please indicate "Submission" in the title of your manuscript email, and indicate your name, phone number, unit and position in the manuscript. Welcome to add my personal WeChat account MooreRen001 or send an email to jyzhang@moore.ren

Click to read the original text and join the Moore Elite