Recently we heard a new Samsung project aims to ‘copy and paste’ a human brain to neuromorphic chips. And now Intel has introduced the Loihi 2, a chip designed for neuromorphic computing.
Loihi 2 brings many improvements over the four-year-old first design Loihi. Almost every element in the chip has been revised, and the bottom line is that the performance should be ten times higher.
While the Loihi was still produced using the 14 nm process, the Loihi 2 is a pre-release version of the upcoming Intel 4 (formerly known as 7+ nm EUV). With 31 mm² instead of 60 mm², the new chip is only about half the size, but still has a million neurons instead of just 128,000.
The internal structure also includes 128 neuron cores in a two-dimensional 8×16 mesh, which, however, contains eight times the number of neurons. Each core has another 128 Kbytes of local SRAM available, but this can now be flexibly partitioned.
In addition, Loihi 2 has gained in flexibility, since it has more extensive microcode, which reinforces the functioning of impulse neurons. This new chip carries a new development framework, called Lava, which is developed in Java and has been released as open-source. Thus, developers can work on programs for Loihi 2 without having access to the hardware. It can be obtained for free from GitHub.
Among other things, Lava integrates a Python interface and can also be used on CPU/GPU platforms, which significantly simplifies development. Frameworks such as Nengo, Pytorch, Robotic Operating System (ROS), Tensorflow or Yarp should also be easy to integrate.
Neuromorphic computing is very different from anything that has been around before, so it is still many years away from series production. The focus is on a very low energy requirement and at the same time high power, microwatt energy levels and millisecond response times are the goal — just like a brain does. Intel did not disclose new performance information for Loihi 2.