Intel Corp. today debuted a new neuromorphic chip, Loihi 2, with 1 million artificial neurons made using a cutting-edge semiconductor manufacturing process not yet in production. 

The company says that Loihi 2 is up to 10 times faster than its previous neuromorphic processor. 

Loihi 2 is significant for several reasons. It represents a new milestone in Intel’s efforts to develop neuromorphic processors, which carry out computations with a vastly different approach than the central processing units for which the company is best known.

Loihi 2 is also notable for the way it’s manufactured. Intel based the chip on a pre-production chip manufacturing process, Intel 4, that’s expected to become fully operational only in the second half of 2022.

“Investigators at Los Alamos National Laboratory have been using the Loihi neuromorphic platform to investigate the trade-offs between quantum and neuromorphic computing, as well as implementing learning processes on-chip,” said Los Alamos National Laboratory staff scientist Gerd Kunde. “We have also demonstrated that the back-propagation algorithm, a foundational building block for training neural networks and previously believed not to be implementable on neuromorphic architectures, can be realized efficiently on Loihi.”

Neuromorphic computing is an emerging approach to semiconductor design focused on developing processors that function like the human brain in certain respects. It’s believed that neuromorphic processors could run artificial intelligence models much faster than traditional chips. Moreover, the technology would require less electricity. 

The growing enterprise adoption of machine learning is driving strong demand for AI chips, as evidenced by the sharp increase in Nvidia Corp.’s share price over recent years. Intel has been actively working to win a bigger share of the market. Though still at an early stage, the company’s neuromorphic chips could eventually give it another way of competing with Nvidia in the AI chip market. 

Loihi 2 is made up of a million computing units dubbed neurons by Intel. The neurons are connected to one another by links called synapses. According to Intel, the synapses allow the computing units to exchange data and coordinate their work. 

At the circuit level, the 1 million neurons in Loihi 2 are made up of 2.3 billion individual transistors. The transistors are fabricated using the Intel 4 manufacturing process, the first from Intel to use a technology known as extreme ultraviolet lithography or EUV. The technology relies on beams of laser light to carve circuits into silicon wafers.

Several other players in the chip market have already adopted EUV, which makes it possible to create more sophisticated transistors than is possible with the earlier manufacturing technology that Intel currently uses. That Intel is now leveraging EUV as well to produce Loihi 2’s transistors sends a positive signal about its efforts to catch up with rivals in chip manufacturing.

One level of abstraction higher, Loihi 2’s 1 million neurons are organized into 128 cores. The cores are connected to each other by a network-on-chip system built into the processor. Compared with Intel’s previous-generation neuromorphic processor, the Loihi 2 provides up to 10 times higher performance with 15 times better hardware density. It also uses less power, the company says.

The chip’s speed and efficiency improvements are attributed to not only the Intel 4 manufacturing process, but also an array of other innovations made by Intel’s engineers. 

The neurons in the company’s neuromorphic chips exchange data with one another through signals dubbed spikes. In Intel’s previous-generation neuromorphic chip, signals could only carry binary data made up of ones and zeros. Loihi 2 adds the ability to send integer numbers in signals, which allows for more sophisticated processing than can be achieved using simple binary values.

Another improvement touted by Intel is increased programmability. Unlike its predecessor, which was optimized for running one specific type of neural network, Loihi 2 can be programmed to run a relatively broad range of neural networks. The chip’s flexibility will enable customers to harness it for a larger number of use cases.

In internal tests, Intel has shown that Loihi 2’s improved design allows it to run the error back-propagation algorithm, an essential component of many AI models. Support for the algorithm represents an important step toward commercialization. Loihi 2 has also been shown to be capable of running inference calculations, the computations that AI models use to analyze data, with up to 60 times fewer operations per inference than Intel’s previous-generation neuromorphic silicon.

Intel detailed Loihi 2 today alongside Lava, an open-source framework designed to help developers build neuromorphic applications for the chip. Lava promises to simplify many of the coding tasks involved in creating software compatible with Loihi 2. Additionally, it enables developers to test neuromorphic applications on standard computing hardware. 

“We are open-sourcing Lava to address the need for software convergence, benchmarking and cross-platform collaboration in the field, and to accelerate our progress toward commercial viability,” said Mike Davies, director of Intel’s Neuromorphic Computing Lab.

Initially, Intel is making Loihi 2 available to customers in two form factors. The first is a system called Oheo Gulch that combines a single Loihi 2 chip with a field programmable gate array from Intel’s Arria 10 product line. The Arria 10 unit allows users to control the neuromorphic chip remotely via an Ethernet connection. Intel expects that the Oheo Gulch will be used mainly in lab projects.

For organizations planning more complex projects, the chipmaker is offering a second system called Kapoho Point, which includes eight Loihi 2 chips. Multiple Kapoho Point units can be linked together to form larger neuromorphic processing clusters. 

Image: Intel