Tiny Chiplets: a New Level of Micro Manufacturing 83
concealment sends this quote from the NY Times:
"Today’s chips are made on large wafers that hold hundreds of fingernail-sized dies, each with the same electronic circuit. The wafers are cut into individual dies and packaged separately, only to be reassembled on printed circuit boards, which may each hold dozens or hundreds of chips. PARC researchers have a very different model in mind. ... they have designed a laser-printer-like machine that will precisely place tens or even hundreds of thousands of chiplets, each no larger than a grain of sand, on a surface in exactly the right location and in the right orientation. The chiplets can be both microprocessors and computer memory as well as the other circuits needed to create complete computers. They can also be analog devices known as microelectromechanical systems, or MEMS, that perform tasks like sensing heat, pressure or motion. The new manufacturing system the PARC researchers envision could be used to build custom computers one at a time, or as part of a 3-D printing system that makes smart objects with computing woven right into them."
It's still silicon (Score:5, Informative)
TFA goes on to say it goes against 50 years of thinking. Spreading out transistors rather than putting them closer together.
They're still placing traditionally produced silicon wafers on what is effectively a printed circuit board. The wafers are just smaller. The method of placing them is new.
I don't see how spreading out parts of a system that operate in the many GHz range is going to help performance. You'll run into problems of electrons not passing charge quick enough because of that pesky speed of light thing. At 3GHz, light only travels 100mm per cycle. Electricity won't go further than 95mm in copper. Half that if you look at the speeds in parts of the core of old P4 processors, upwards of 7GHz.
Here's the patent application (Score:5, Informative)
It has the same components as a traditional Xerox machine. There is a drum that rotates and their positioning technology put the chiplets in precise locations on the substrate. The chiplets are in a fluid that acts like toner.
It appears that the performance depends on how fast the substrate conducts signals. At this point it seems unlikely that this is as fast as an on chip connection, but there seems to be no intrinsic reason that it would be any slower then the wires that hook a chip pad to a package pin. In aggregate the speed might be faster then a circuit board because the chiplits could be closer together then chips on a board.
One possible deployment would be to use this to assemble components which are then packaged in a standard IC. It's like an SOIC, except the parts are not all on one piece of silicon.
There are potential economies of scale. With an inventory of chiplets, and automation to make the interconnect substrate with CAD, a custom assembly line can create vast numbers of different configurations and not have to include a foundry in the loop.
Despite all the naysayers that have already posted, this is a potentially game changing technology.