AI systems have a brawny new chip for the device edge: Intel’s Pohoiki Springs

Categories Edge Computing News  |  Hardware

Intel-Pohoiki-Springs - Intel Loihi neuromorphic research chips
(Intel Loihi neuromorphic research chips. Tim Herman/Intel Corporation)

Three years after Intel Corp. introduced its Loihi neuromorphic computer chips, the company has chained 768 of them into a chassis providing the computational capacity of 100 million neurons.

A hamster has that many natural neurons. In this case, Intel is making the equivalent of a rodent brain for edge computing, where ultra-fast operations and ultra-low power use are most prized.

The newly integrated chips are part of a new cloud-based research system that Intel calls Pohoiki Springs, which is being made available to members of Intel’s neuromorphic research community. The group was formed two years ago to experiment with what was then a test version of the Loihi chip.

In announcing the Pohoiki Springs, Intel executives said the chip operates at less than 500 watts. Meanwhile, Kapoho Bay, Intel’s smallest neuromorphic system, uses two Loihi chips to support real-time edge workloads like recognizing gestures in real-time and orienting direction using learned visual landmarks. The chips consuming only tens of milliwatts of power, a key measure for edge use cases that require hardware with low power needs.

Probably because the Pohoiki Springs is aimed at an internal research audience rather than commercial production, no price has been announced for the system.

Company executives describe Loihi chips as being up to 1,000 times faster and 10,000 times as efficient performing certain functions. The integrated processors exhibit extreme parallelism and asynchronous systems, which account for the performance improvements.

Mike Davies, director of Intel’s Neuromorphic Computing Lab, was quoted in electronic-engineering publication EETimes, said the Pohoiki Springs does not work the same way artificial intelligence systems do, and it is hard to compare them.

Neuromorphic systems function in a way not unlike how living brains function. As such, they are better for different tasks than one generally would feed to artificial-intelligence algorithms.

The systems process individual real-world data samples immediately with low latency and power consumption, Davies said.

He said this method is well-designed for edge environments because it enables models that can adapt and learn from the real-time samples.

Intel said researchers are working on “promising, highly scalable algorithms” to take advantage of the new architecture. They include constraint satisfaction, searching graphs and patterns and optimization problems.

The smallest neuromorphic system Intel has, the edge-focused Kapoho Bay, has 262,000 neurons, which is on par with the brain of an ant.

Article Topics

 |   |   | 


Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Edge Analysis & Opinion

Featured Company

Edge Market Reports

Explaining Edge Computing

Edge Computing Events

Latest News