Slashdot Mirror
DARPA Has an Ambitious $1.5 Billion Plan To Reinvent Electronics (technologyreview.com)
The Defense Advanced Research Projects Agency (DARPA), which funds a range of blue-sky research efforts relevant to the US military, last year launched a $1.5 billion, five-year program known as the Electronics Resurgence Initiative (ERI) to support work on advances in chip technology. It has now unveiled the first set of research teams selected to explore unproven but potentially powerful approaches that could revolutionize US chip development and manufacturing. From a report: The ERI's budget represents around a fourfold increase in DARPA's typical annual spending on hardware. Initial projects reflect the initiative's three broad areas of focus: chip design, architecture, and materials and integration. One project aims to radically reduce the time it takes to create a new chip design, from years or months to just a day, by automating the process with machine learning and other tools so that even relatively inexperienced users can create high-quality designs.
"No one yet knows how to get a new chip design completed in 24 hours safely without human intervention," says Andrew Kahng of the University of California, San Diego, who's leading one of the teams involved. "This is a fundamentally new approach we're developing." William Chappell, the head of the DARPA office that manages the ERI program, said, "We're trying to engineer the craft brewing revolution in electronics." The agency hopes that the automated design tools will inspire smaller companies without the resources of giant chip makers, just as specialized brewers in the US have innovated alongside the beer industry's giants.
"No one yet knows how to get a new chip design completed in 24 hours safely without human intervention," says Andrew Kahng of the University of California, San Diego, who's leading one of the teams involved. "This is a fundamentally new approach we're developing." William Chappell, the head of the DARPA office that manages the ERI program, said, "We're trying to engineer the craft brewing revolution in electronics." The agency hopes that the automated design tools will inspire smaller companies without the resources of giant chip makers, just as specialized brewers in the US have innovated alongside the beer industry's giants.
I'm beginning to think that we're reaching the limits of what we can do with the laser lithography method of silicon IC creation. For instance look at the problems Intel is having with 10nm fabrication right now. Perhaps the way forward is straight out of science fiction: a matter compiler/3D printer-like approach, where an integrated circuit is built up an atom or a molecule at a time? Pure imagination on my part, but is it really out of our reach?
eliminating the thousands or tens of thousands of timing violations. ASIC development goes a compilation process of source code not entirely different than software. The challenge of ASIC design is locating everything just right so that distance is globally nearly-minimized (local optima but near the global optimum) to get all the electron pulses to arrive where they need to arrive before it is too late. Plus, each logic gate costs a time delay (as well as occupies space exacerbating the distance problem). So another trick to solve timing violations is to simplify the design is some locality to lessen the depth of gates that a signal/calculation/operation must traverse, when viewed as a directed-acyclic graph (DAG). Sometimes space (# of gates) can be bloated up to decrease the depth of the walks of the logic-gate DAG (but then that increases area on the die, which exacerbates the distance problem).
Not to mention that the tools that area already automated to *assist* with this stuff are buggy as fuck, and EDA companies software development practices make Microsoft look good.
But more power to DARPA, the more that can be automated, the more that can be accomplished. Chip design is still so expensive that only a few people with very deep pockets can participate.
FPGAs have advanced a lot in 20 years, but they are the moral equivalent of running your code under Java. That's not at all what they do, but their ability to configure on the fly comes at a very high cost in terms of frequency, power and area. A custom chip is always going to be faster, smaller and cheaper (COGS wise).
It's true that FPGA mfg's could do more to enable other tools, but their motivation is very weak.
Given all the design work is done on the workstation by typing a lot of HDL code and doing spice simulations and maybe even some layout work, the particular focus of this one is purely a software investment.
I don't know but obviously it's bullshit; I didn't see any mention of "AI" anywhere.
Basic research! Nothing more likely to "fail" in the commercial sense, and so hated by free market companies that hate risk (all of them); nothing more likely - longer term - to come up with the big finds that create whole new economic sectors.
Their "ARPAnet" idea wasn't even supposed to make money, that's the funny bit.
With just a little luck, some of this research will end up creating whole new economic opportunities, which will result in a few people becoming billionaires, who will probably, with tiresome regularity, turn out to be libertarians who don't believe government can do anything useful and attempt to pay no taxes.
Ah, those public bureaucrat-scientists struggling for grants: America's true Job Creators.
(Juuuust kidding, of course. America's real job creators are consumers: without people putting butts in seats of the restaurant, neither the cooks&waiters, nor the restaurant owner, nor his banker, have any jobs.)