The vast majority of chips for artificial intelligence that we currently have in our PCs are designed to execute inference algorithms, this means that what they do is make predictions or classifications from a “deductive” method that has been previously trained in another system and many times several times more powerful than the hardware we have at home.
Due to this, today one of the fields in which we seek to advance more quickly is in machine learning, because the reason for this is that the methods used require huge data sets for learning. Making a simile with a human being is as if we will need something to be explained many times to understand it.
The solution by computer scientists? Create processors whose operation resembles biological neural structures as much as possible, since after all it is what we know in nature and have the demonstrated ability in all species to learn on the fly.
Intel Loihi 2, the second generation of Intel’s neuromorphic processor
The largest manufacturer of PC CPUs in the world not only focuses its research and development on creating more powerful CPUs, but a company of this size given the continuous advances in the different fields of computing cannot afford to stand still. And that is why in 2018 they created the Intel NRC laboratory, Neuromorphic Research Community, from where they designed the first Loihi. Now, three years later they have presented their improved second generation.
As you can see in the images, Loihi 2 is a very small chip, since it measures only 31 square millimeters and it is the first chip built under the Intel node 4, its node equivalent to the 5 nm of TSMC, which is the same that Intel will use for the creation of Meteor Lake, its CPU for PC Intel Core Gen 14 that we will see in late 2023. Although with respect to its architecture this neuromorphic processor does not have nothing to do with that future CPU and rather it serves to demonstrate that it can already manufacture processors under said manufacturing process.
Domain-specific and 3D processors to accelerate machine learning
Regarding the complexity of the chip, it is made up of a 8 x 16 processor 2D matrix, which is the typical interconnection of the cores for AI, each of them has its own 128 KB local memory and it is interconnected with the rest of the nuclei forming a NoCTherefore, although Intel has not said it, we deduce that each core has a router that allows it to communicate with its surroundings directly.
The processors in Loihi 2 are very straightforward, measuring just barely 0.44 square millimeters and it has a very simple set of instructions, as you can see in the table on these lines. So we are not dealing with cores as complex as the ones we use in our PCs, they are what we call domain-specific processors, which are programmable, but only have what is necessary for the task they have been designed for.
In addition, Loihi 2 has been designed to scale in number of cores making use of the technology of vertical interconnection from Intel, Foveros, so multiple Loihi 2 chips can be placed to obtain more complex versions of it. This capacity is also due to the fact that it is a chip not only very small, but also very low consumption, 1 watt only, so energy consumption and generated heat are not a problem for the creation of the first neuromorphic chip in the form of an integrated circuit that can scale vertically and therefore be the first AI chip in 3DIC.
Lava, the Loihi 2 programming environment
Every processor needs programs to run and developers need an environment to write them, and this is where Lava comes in, which is an open source framework that Intel has designed and that allows the development of artificial intelligence algorithms written in programming languages. like Python and C.
One of the key elements of Lava is Magma, which is a low-level interface that allows the deployment of artificial intelligence algorithms that have been previously developed in a CPU or GPU, but which are intended to be executed in Loihi 2. Furthermore, Lava is designed to be integrated into third-party frameworks and for ease of use all the libraries that Intel offers as standard are written in Python.
What are the potential markets for this neuromorphic processor?
There are a number of fields where Intel has been testing Loihi 2 in its NRC lab, but one of the most common applications for this type of processor is computer vision, which should not be confused with computer graphics. While what a GPU does is draw and display an image, computer vision copies the mechanisms of nature for the vision and processing of images and therefore of one of the senses. And in the same way that we have computer vision we can also apply it to other classic senses such as the detection of odors, interpretation of sounds and even touch.
This is important due to the fact that in self-learning, artificial intelligence systems do not start from a set of previous information that has been ordered and classified, but rather have to capture information from the environment in real time as input data for inference. . And since we base artificial intelligence on nature-based systems, we also have to give them the ability to see, hear, feel and even smell.
At a commercial level, the applications of this type of chips are various, from the well-known automatic driving of military vehicles or drones to automated production plants and even assistance systems in surgery. In particular, what is sought is to have systems with the ability to learn by themselves from their mistakes.