US Focuses on Invigorating 'Chiplet' Production in the US (nytimes.com) 19
More than a decade ago engineers at AMD "began toying with a radical idea," remembers the New York Times. Instead of designing one big microprocessor, they "conceived of creating one from smaller chips that would be packaged tightly together to work like one electronic brain."
But with "diminishing returns" from Moore's Law, packaging smaller chips suddenly becomes more important. [Alternate URL here.] As much as 80% of microprocessors will be using these designs by 2027, according to an estimate from the market research firm Yole Group cited by the Times: The concept, sometimes called chiplets, caught on in a big way, with AMD, Apple, Amazon, Tesla, IBM and Intel introducing such products. Chiplets rapidly gained traction because smaller chips are cheaper to make, while bundles of them can top the performance of any single slice of silicon. The strategy, based on advanced packaging technology, has since become an essential tool to enabling progress in semiconductors. And it represents one of the biggest shifts in years for an industry that drives innovations in fields like artificial intelligence, self-driving cars and military hardware. "Packaging is where the action is going to be," said Subramanian Iyer, a professor of electrical and computer engineering at the University of California, Los Angeles, who helped pioneer the chiplet concept. "It's happening because there is actually no other way."
The catch is that such packaging, like making chips themselves, is overwhelmingly dominated by companies in Asia. Although the United States accounts for around 12 percent of global semiconductor production, American companies provide just 3 percent of chip packaging, according to IPC, a trade association. That issue has now landed chiplets in the middle of U.S. industrial policymaking. The CHIPS Act, a $52 billion subsidy package that passed last summer, was seen as President Biden's move to reinvigorate domestic chip making by providing money to build more sophisticated factories called "fabs." But part of it was also aimed at stoking advanced packaging factories in the United States to capture more of that essential process... The Commerce Department is now accepting applications for manufacturing grants from the CHIPS Act, including for chip packaging factories. It is also allocating funding to a research program specifically on advanced packaging...
Some chip packaging companies are moving quickly for the funding. One is Integra Technologies in Wichita, Kan., which announced plans for a $1.8 billion expansion there but said that was contingent on receiving federal subsidies. Amkor Technology, an Arizona packaging service that has most of its operations in Asia, also said it was talking to customers and government officials about a U.S. production presence... Packaging services still need others to supply the substrates that chiplets require to connect to circuit boards and one another... But the United States has no major makers of those substrates, which are primarily produced in Asia and evolved from technologies used in manufacturing circuit boards. Many U.S. companies have also left that business, another worry that industry groups hope will spur federal funding to help board suppliers start making substrates.
In March, Mr. Biden issued a determination that advanced packaging and domestic circuit board production were essential for national security, and announced $50 million in Defense Production Act funding for American and Canadian companies in those fields. Even with such subsidies, assembling all the elements required to reduce U.S. dependence on Asian companies "is a huge challenge," said Andreas Olofsson, who ran a Defense Department research effort in the field before founding a packaging start-up called Zero ASIC. "You don't have suppliers. You don't have a work force. You don't have equipment. You have to sort of start from scratch."
But with "diminishing returns" from Moore's Law, packaging smaller chips suddenly becomes more important. [Alternate URL here.] As much as 80% of microprocessors will be using these designs by 2027, according to an estimate from the market research firm Yole Group cited by the Times: The concept, sometimes called chiplets, caught on in a big way, with AMD, Apple, Amazon, Tesla, IBM and Intel introducing such products. Chiplets rapidly gained traction because smaller chips are cheaper to make, while bundles of them can top the performance of any single slice of silicon. The strategy, based on advanced packaging technology, has since become an essential tool to enabling progress in semiconductors. And it represents one of the biggest shifts in years for an industry that drives innovations in fields like artificial intelligence, self-driving cars and military hardware. "Packaging is where the action is going to be," said Subramanian Iyer, a professor of electrical and computer engineering at the University of California, Los Angeles, who helped pioneer the chiplet concept. "It's happening because there is actually no other way."
The catch is that such packaging, like making chips themselves, is overwhelmingly dominated by companies in Asia. Although the United States accounts for around 12 percent of global semiconductor production, American companies provide just 3 percent of chip packaging, according to IPC, a trade association. That issue has now landed chiplets in the middle of U.S. industrial policymaking. The CHIPS Act, a $52 billion subsidy package that passed last summer, was seen as President Biden's move to reinvigorate domestic chip making by providing money to build more sophisticated factories called "fabs." But part of it was also aimed at stoking advanced packaging factories in the United States to capture more of that essential process... The Commerce Department is now accepting applications for manufacturing grants from the CHIPS Act, including for chip packaging factories. It is also allocating funding to a research program specifically on advanced packaging...
Some chip packaging companies are moving quickly for the funding. One is Integra Technologies in Wichita, Kan., which announced plans for a $1.8 billion expansion there but said that was contingent on receiving federal subsidies. Amkor Technology, an Arizona packaging service that has most of its operations in Asia, also said it was talking to customers and government officials about a U.S. production presence... Packaging services still need others to supply the substrates that chiplets require to connect to circuit boards and one another... But the United States has no major makers of those substrates, which are primarily produced in Asia and evolved from technologies used in manufacturing circuit boards. Many U.S. companies have also left that business, another worry that industry groups hope will spur federal funding to help board suppliers start making substrates.
In March, Mr. Biden issued a determination that advanced packaging and domestic circuit board production were essential for national security, and announced $50 million in Defense Production Act funding for American and Canadian companies in those fields. Even with such subsidies, assembling all the elements required to reduce U.S. dependence on Asian companies "is a huge challenge," said Andreas Olofsson, who ran a Defense Department research effort in the field before founding a packaging start-up called Zero ASIC. "You don't have suppliers. You don't have a work force. You don't have equipment. You have to sort of start from scratch."
Has somebody (Score:2)
already trademarked the term chiplet?
Wow chips on substrates! (Score:3)
Reminds me of old times, things were roaring along, I am back in the semiconductor fab, some guy named Reagan was president.
Re:Wow chips on substrates! (Score:5, Insightful)
It's more complex than that. For a CPU the chiplets need to be interconnected at speeds that were impossible in the 80s. This isn't like what Apple does with on substrate RAM, it's parts of a multicore CPU.
The performance speaks for itself. Battery life rivalling ARM, performance far exceeding it. It's cheap too - AMD can fabricate different parts of the CPU on different process nodes to concentrate the highest end manufacturing where it matters most. Yields way up too.
Re: (Score:3)
The interconnects are really important though, and it's another area where Intel is struggling. Even if they do manage to overtake TSMC and regain a node advantage, I don't think they'll be able to do a lot with it.
If anything I wish AMD would be able to figu
Re: (Score:2)
Intel only manage passable battery life because they are now building CPUs with efficiency cores. The same is true of ARM, including Apple's M1 and M2. They all use a mixture of performance and efficiency cores to reach low power levels.
In contrast, mobile Ryzen parts are all performance cores. The chiplet design helps them scale the voltage and frequency over wider ranges, and get 20+ hours of battery life on some laptops. It's really impressive stuff.
Re: (Score:2)
The problem Intel has is that their regular cores are too large for them to put 16 of them on a die to go head to head with AMD's Zen cores. AMD has a leaner design that performs just as well with significantly few
This was being talked about in the (Score:4, Insightful)
Re: (Score:2)
Re: (Score:2)
DEC in the early to mid 1980s.
https://en.wikipedia.org/wiki/... [wikipedia.org]
Going out on a limb (Score:1)
I think it's a safe bet that Biden has zero to no clue what this is all about let alone what the money in the CHIPS act is ultimately going to be spent on.
They'll also be trying to patent base 2 and above (Score:2)
In Addition... (Score:2)
It would be really nice to standardize some interchangeable modules so we could end up with solid state components similar to vacuum tubes such that you could replace small, standardized pieces with similar pieces, so there wouldn't be these single-purpose chips that, once they go out of production, it is impossible to repair a $2000 radio of 20 years ago (Like my Kenwood TS-2000) if one of the unique chips goes belly up. My TS-2000 is fine, but lots of other folks have boat anchors that look like my TS-20
Re: (Score:2)
The speeds involved kinda make the "module" idea impossible. The interconnect winds up being far larger and expensive than everything else. High speed connectors that can reliably plug and unplug many times are really difficult to make.
The $2000 radio of 20 years ago would have been much more expensive, larger, and power hungry, if it was built with interchangeable parts.
Re: (Score:2)
High speed interconnects are often rated around... 20 plug-unplug cycles. Yes, 20. They typically last a bit more, but reliability goes out the door beyond 30 or so connections.
And sockets can be even more expensive - sockets for SoCs and stuff costed around... $400 each. And
Re: (Score:2)
Like these?
https://www.pcmag.com/encyclop... [pcmag.com]
Re: (Score:1)
Yeah, kinda like that but not so granular. Integrated circuits of standardized functions, capable of way more than discrete transistors, but engineered to be interchangeable like vacuum tubes. Your golly-gee-whiz radio goes tits up, you can start by pulling every standardized IC in the thing, taking them to the drug store, and inserting them 1 by 1 into the tester until you find the bad one. The nice man at the store retrieves a replacement for you from the drawers of IC's inside the machine, you go ho
Re: (Score:2)
Transistors are the replacement for vacuum tubes, and those little discrete ones are drop in replaceable and made by dozens of manufacturers for a few cents per, compared to tubes that started around $20 in today's money.
You can also buy ICs of course. Thousands of different standard ones, made by many manufacturers, usually for less than a dollar each. Radio Shack used to sell them, and there are still some chips-in-drawers retailers, but most people use Amazon or Aliexpress these days.
Neither the transist