Over the past decade, AI has gone from theory and small-scale testing to enterprise-scale use.
New research from the IBM Research lab (focused on human brain model computing) in Almaden, California has the potential to fundamentally change the approach to effectively scale up powerful AI hardware systems.
For the past 8 years, Modha, the lab's lead researcher, has been working on a new type of digital AI chip for neural computing. The chip, dubbed NorthPole, is an extension of TrueNorth, the last chip that Modha worked on until 2014.
According to Modha, NorthPole is a breakthrough in chip architecture, providing significant improvements in energy, spatial and temporal efficiency. NorthPole is manufactured using a 12-nanometer process and contains 22 billion transistors in an 800-square-millimeter area. It has 256 cores and is capable of performing 2,048 operations per core per cycle at 8-bit precision with the ability to double and quadruple the number of operations at 4- and 2-bit precision, respectively. "It's an entire network on a single chip," Modha said.
NorthPole's biggest advantage is simultaneously a limitation: it can only utilize the memory it has on board. All the acceleration possible on the chip would be negated if it had to access information from elsewhere. Through an approach called scaling, NorthPole can support large neural networks by breaking them up into smaller subnetworks that fit into the memory of the NorthPole model, and connecting these subnetworks to each other across multiple NorthPole chips.
For the past 8 years, Modha, the lab's lead researcher, has been working on a new type of digital AI chip for neural computing. The chip, dubbed NorthPole, is an extension of TrueNorth, the last chip that Modha worked on until 2014.
According to Modha, NorthPole is a breakthrough in chip architecture, providing significant improvements in energy, spatial and temporal efficiency. NorthPole is manufactured using a 12-nanometer process and contains 22 billion transistors in an 800-square-millimeter area. It has 256 cores and is capable of performing 2,048 operations per core per cycle at 8-bit precision with the ability to double and quadruple the number of operations at 4- and 2-bit precision, respectively. "It's an entire network on a single chip," Modha said.
NorthPole's biggest advantage is simultaneously a limitation: it can only utilize the memory it has on board. All the acceleration possible on the chip would be negated if it had to access information from elsewhere. Through an approach called scaling, NorthPole can support large neural networks by breaking them up into smaller subnetworks that fit into the memory of the NorthPole model, and connecting these subnetworks to each other across multiple NorthPole chips.