[HOT] : Intel Announces Second Generation of Loihi Neuromorphic Chips

On Thursday, Intel announced the second version of the Loihi neuromorphic chip, the Loihi 2. It is an artificial intelligence processor designed to more closely mirror the processes that occur in the human brain than other technologies in the world. IA. The founder says the new chip is now halved with 31mm2, more advanced process nodes. It contains 1 million artificial spiked neurons, eight times its predecessor.

The chip becomes more flexible thanks to the larger microcode that underlies the functioning of the pulsed neurons. Pronounced “low-EE-he”, Loihi takes its name from the underwater volcano in Hawaii, which “emerges from the ocean,” as Mike Davies, director of neuromorphic computing at Intel, explains. .. The new chip also comes with a new development framework called Lava, written in Java and released as open source. Therefore, developers can work with Loihi programs without having to access hardware. This software is available on Github.

As a security measure, Intel announced the first version of Loihi in 2017. At the time, it was described as a chip “inspired by the way neurons communicate and learn using timely adjustable spikes and plastic synapses ”. The chip is designed as a “mesh” of computing cores and praised by the founders of the United States.

More flexibility for programmers

The first Loihi was designed using Intel’s 14 nanometer processing technology. The new version uses Intel 4 technology. This is because the company has renamed the latest generation of manufacturing. Intel isn’t disclosing all the details of the processing technology, but it’s basically a 7-nanometer chip. The technology is a prototype version of Intel 4, the company points out.

The chip will benefit from more flexible microcode that allows programmers to “assign variables and execute a wide range of instructions organized in the form of short programs using assembly language,” said Intel. These programs have access to the neural state memory, to the accumulated synaptic inputs, to the random bits of the probabilistic model and to the time step counters to be calculated as a function of time. The instruction set supports conditional branching, bitwise logic, and fixed-point arithmetic supported by hardware multipliers.

“This new chip adds a lot of generality and programmability, but was previously limited by Loihi 1’s fixed function set,” says Mike Davies, who was interviewed by ZDNet. He added that Intel has used chips to improve performance on several levels. “I have workloads that run over 10 times faster, and at the circuit level it’s between 2 and 7 times, depending on the configuration. Intel also quadrupled the bandwidth of each link between the cores. In doing so, “the scalability between chips” has been improved.

Criticized neuromorphic approach

According to Mike Davies, the chips can be scaled in three dimensions using inter-chip links. Proponents of neuromorphic computing argue that this approach more closely mimics the actual characteristics of brain functioning and the economy of the brain sending signals.

However, deep learning players who take a different approach to AI, unlike deep learning systems like ResNet, which can find images in photos, take the neuromorphic approach because they don’t have obtained practical results. In 2019, Facebook’s chief AI officer, Yann LeCun, rejected the neuromorphic approach during a conference also given by Mike Davies.

However, an article published earlier this year on IEEE Spectrum provided evidence that IEEE Spectrum is being touted as a demonstrable benefit of neuromorphic chips for a variety of problems. He says neuromorphic chips have advantages when there are issues like recursion related to how the brain processes data.

“For the first time, we have a portfolio of results that quantitatively confirm the potential of neuromorphic hardware to deliver significant improvements in energy efficiency, processing latency and data efficiency for certain learning algorithms. I have it, ”he said.

“Great” tip for optimization

“We were delighted and surprised to learn that neuromorphic chips are ideal for solving optimization problems,” says Mike Davis. “Rather, forms of deep learning are strictly ‘feedforward’ in prediction and reasoning, even though the brain is a much more complex process that constantly considers feedback, understanding the context, and expectations within it. Defend the process. Inference, ”he emphasizes.

One of the first users of Loihi 2, the famous Los Alamos National Laboratory of the US Department of Energy, said it was able to use a chip to connect between multiple domains. The use of Loihi has highlighted a crossroads in machine learning and quantum computing, said researcher Gerd J. Kunde.

“Our research has shown that backpropagation algorithms, a fundamental building block of neural network learning previously believed to be impossible to implement in neuromorphic architectures, can be performed efficiently in Loihi.” -he declares.

Tagged in :

, , ,