Water-based chips may very well be the way forward for neural networks
Researchers are engaged on water-based microprocessors that might someday be used as a extra numerous different to the present wafer structure of at present, with functions starting from AI to DNA synthesis and certain past.
The chips in query are nonetheless within the prototype stage, so do not count on processors with in-built water cooling simply but, however the best way they work is basically thrilling. They use a method known as ionics, which entails manipulating completely different ion species in liquid, versus the usual electrons capturing via our semiconductors at present.
It is extra akin to how the human mind handles information processing, and though the method could also be slower than present semiconductor architectures, scientists imagine “the variety of ionic species with completely different bodily and chemical properties may very well be harnessed for richer and extra numerous info processing.”
That is in line with a submit (opens in new tab) from the Harvard John A. Paulson College of Engineering and Utilized Sciences (SEAS). Collaborating with startup DNA Script (opens in new tab), researchers from SEAS just lately printed findings describing how they mixed ionic diodes and transistors—one thing no analysis group had accomplished earlier than—to develop an ionic circuit.
Though the easy 16 x 16 array is a great distance from matching the ability of the 100 billion transistors present in Intel’s twelfth Gen wafers, or the Apple M1 Extremely’s 114 billion transistors, the researchers did handle to carry out a neural internet computing core course of.
They used mentioned array to “increase the analog arithmetic multiplication of particular person transistors into an analog matrix multiplication.” Yeah I am fairly misplaced right here, too. Fortunately SEAS postdoctoral fellow Woo-Bin Jung explains that “Matrix multiplication is essentially the most prevalent calculation in neural networks for synthetic intelligence.” He says their “ionic circuit performs the matrix multiplication in water in an analog method that’s primarily based totally on electrochemical equipment.”
The transistors work via manipulation of native PH values, which permits the researchers to manage the chips’ weight matrix, equally to that of a neural community.
Not solely may there be potential for functions comparable to accelerating DNA synthesis (opens in new tab) and different neural networking processes, Professor of Electrical Engineering and Utilized Physics at SEAS Donhee Ham says “the electrochemical matrix multiplication in water is charming in its personal proper, and has a possible to be vitality environment friendly.”
To date, Jung notes that solely “3 to 4 ionic species, comparable to hydrogen and quinone ions” have been examined, however because the researchers transfer on to testing different ionic species, the knowledge processing is just going to grow to be richer and extra numerous.
Perhaps someday we’ll see neural networks operating off water-based ionic chips. They’re going to be way more vitality environment friendly, however possible a lot slower.