Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/google%E2%80%99s-ai-approach-to-microchips-is-welcome-%E2%80%94-but-needs-care.14337/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2020570
            [XFI] => 1050070
        )

    [wordpress] => /var/www/html
)

Google’s AI approach to microchips is welcome — but needs care

Daniel Nenni

Admin
Staff member
I think this is more ML than AI but progress just the same. I feel another semiconductor disruption coming, absolutely.


Artificial intelligence can help the electronics industry to speed up chip design. But the gains must be shared equitably. Google’s researchers used 10,000 chip floorplans to train their software. The software then worked out how to generate floorplans that used no more space, wire and electric power than did those designed by engineers. Miniaturization and low power are particularly important for the chips used in smartphones.

The AI-generated chips took less than six hours to design, and the method has already been used to design Google’s tensor processing unit, or TPU, which is used mainly in the company’s cloud-based machine-learning applications. More teams need to test the design software to make sure it is robust and can accommodate other data sets and chip types. If more groups can recreate its success, that will cement its place in the chip-design toolbox.
 
Chips are already designed by application programs that use
optimization algorithms from operations research. There have
been questionable claims of algorithm improvements going
back to the Timber Wolf layout program. Google is suggesting
oblivious neural network optimization programs. Ther is volumunous
human expert knowledge is current layout programs that neural
networks are claimed to not need.
 
They couldn't do anything worse than the current EDA cartels. Google made their bones on open source software, and their experience with the TPUs and their efforts with Skywater 130nm hopefully informed them that the closed source tool chain is the thing inhibiting grassroots innovation. Say what you want about their spooky management, at least their engineers are still tech idealists who believe in open competition.
 
Top