Intel Unveils Loihi 2 Neuromorphic Processor, And Samsung – Copy And Paste Concept Of Brain Neural Connections On Neuromorphic Chips | Gadgets News

Today, four years after the announcement of Loihi, Intel introduced the 2nd generation of this neuromorphic processor. Here’s what the characteristics of both chips look like:

Loihi 2 Long
Technical process Intel 4 (former 7nm) Intel 14 nm
Crystal area 31 mm 2 60 mm 2
Number of transistors 2.3 billion 2.1 billion
Number of neurons 1 million 128 thousand
Quantity in the synapse 120 million 128 million
Number of cores 128 128
Single core cache 208 Kb 192 Kb

The company says that in operations with deep neural networks, the performance of the Loihi 2 is 10 times higher than that of its predecessor.

As Gadgets News has already told, in the spring of last year, Intel introduced the Pohoiki Spring computing cluster of 768 Loihi chips and, accordingly, 98.3 million neurons. Thus, a similar system from Loihi 2 (for which Intel has not yet disclosed its plans) will contain 768 million neurons.

For comparison, the human brain contains, according to various estimates, 80-90 billion neurons (about a fifth of which are located in the cerebral cortex), and here we smoothly turn to another interesting news. The other day Samsung presented “A new concept that brings the world one step closer to creating neuromorphic chips that allow better reproduction of the brain.”… Here’s how it is stated in the company’s press release:

The idea, proposed by leading engineers and scientists from Samsung and Harvard University, was published by Nature Electronics as a review article entitled “Neuromorphic Electronics Based on the Copy and Paste of Neural Connections in the Brain.”

The essence of the concept put forward by the authors is best expressed by the two words “copy” and “paste”. This article proposes a way to copy a map of brain neural connections using the revolutionary nanoelectrode array developed by Dr. Ham and Dr. Park, and insert this map into a high-density 3D solid-state memory network, in whose development technology Samsung is the world leader.

Using this copy-and-paste approach, the authors propose to create a memory microcircuit that, in terms of its characteristics, will approach the unique computing capabilities of the brain – it will have low power consumption, be easy to learn, adapt to the environment, and even be autonomous and support cognitive functions – all of this. is beyond the reach of current technology.

The brain consists of a large number of neurons, and the various connections between them are responsible for the functions of the brain. Thus, knowing the map of these connections is the key to reverse engineering the way the brain works.

Although the original goal of neuromorphic engineering, which dates back to the 1980s, was to replicate the structure and operation of neural networks on a silicon chip, this task has proven extremely difficult – even now, scientists know little about what connections exist. between a large number of neurons involved in higher brain functions. As a result, the goal of neuromorphic engineering was simplified and reduced to the creation of a chip, more “based on” the brain, rather than an exact repetition of it.

This article proposes a way to go back to the original neuromorphic goal of brain reverse engineering. The array of nanoelectrodes, in fact, can enter a large number of neurons, due to which it is able to record their electrical signals with high sensitivity. These massively parallel intracellular recordings allow you to create a map of neural connections, indicating the connection points of neurons and the strength of these connections. Thus, a map of neural connections can be extracted or “copied” from these signaling records.

The copied neural map can then be “inserted” into a nonvolatile memory network – for example, commercial flash memory, which is used in our daily life in solid-state drives (SSD), or into “new” memory, for example, resistive random access (resistive random access memories, RRAM) – by programming each memory element so that its conductance represents the strength of each neural connection in the copied map.

The authors of the article develop this idea and propose a strategy for quickly inserting the resulting map of neural connections into the memory network. A network of specially designed non-volatile memory devices can study and reproduce a map of neural connections under direct control from signals recorded at the intracellular level. Basically, such a circuit directly loads a map of the brain’s neural connections into a memory chip.

Since the human brain has about 100 billion neurons and about a thousand times more synaptic connections, it will take about 100 trillion memory elements to create a neuromorphic chip. Combining such a huge amount of memory elements on a single chip is made possible by 3D memory integration, a technology developed by Samsung that will usher in a new era in the memory industry.

Building on its best practices in chip manufacturing, Samsung intends to continue research in neuromorphic engineering to strengthen Samsung’s leadership in next-generation semiconductors for artificial intelligence technologies.

“The idea we presented is very bold and ambitious,” said Dr. Ham. “Our work and progress towards such an ambitious goal will allow us to push the boundaries of machine intelligence, neuroscience and semiconductor technology.”

We would like to say thanks to the author of this post for this awesome content

Intel Unveils Loihi 2 Neuromorphic Processor, And Samsung – Copy And Paste Concept Of Brain Neural Connections On Neuromorphic Chips | Gadgets News