Ibm Says It Has An All-In-One Deep-Learning Chip

From IEEE Spectrum:

IBM’s New Do-It-All Deep Learning Chip
IBM's novel chip is designed to produce both high-precision learning in addition to low-precision inference across the 3 principal flavors of deep learning

The patch of deep learning is nonetheless inwards flux, exactly approximately things convey started to settle out. In particular, experts recognize that neural nets tin acquire a lot of computation done amongst footling energy if a chip approximates an response using low-precision math. That’s particularly useful inwards mobile in addition to other power-constrained devices. But approximately tasks, particularly preparation a neural internet to produce something, nonetheless demand precision. IBM lately revealed its newest solution, nonetheless a prototype, at the IEEE VLSI Symposia: a chip that does both as well.

The disconnect betwixt the needs of preparation a neural internet in addition to having that internet execute its function, called inference, has been i of the large challenges for those designing chips that accelerate AI functions. IBM’s novel AI accelerator chip is capable of what the fellowship calls scaled precision. That is, it tin produce both preparation in addition to inference at 32-, 16-, or fifty-fifty 1- or 2-bits.

“The almost advanced precision that you lot tin produce for preparation is sixteen bits, in addition to the almost advanced you lot tin produce for inference is 2 bits,” explains Kailash Gopalakrishnan, the distinguished fellow member of the technical staff at IBM’s Yorktown Heights inquiry middle who led the effort. “This chip potentially covers the best of preparation known today in addition to the best of inference known today.”

The chip’s powerfulness to produce all of this stems from ii innovations that are both aimed at the same outcome—keeping all the processor components fed amongst information in addition to working.

“One of the challenges that you lot convey amongst traditional [chip] architectures when it comes to deep learning is that the utilization is typically real low,” says Gopalakrishnan. That is, fifty-fifty though a chip mightiness endure capable of a real high superlative performance, typically alone xx to thirty percentage of its resources tin actually endure brought to touching on a problem. IBM aimed for ninety percent, for all tasks, all the time....MORE