Intel introduced Loihi 2, a new version of its neural processor that can simulate neurons and synapses and is therefore somewhat self-learning. The company uses Intel 4 as a test process to make the chip.
According to Intel The basic architecture of Loihi 2 is the same as that of the first Loihi. This neural chip was made by Intel in 2017. Intel makes it at 14nm, with the Loihi 2 being produced on Intel 4, or Intel’s upcoming 7nm process. At the moment, this relates to the pre-production capacity that Intel is using since the Intel 4 is not yet production ready.
Despite the same basic architecture, according to Intel, there are the necessary improvements. Using a smaller node means that the slice cores occupy a smaller surface. The limit is still 128 neurons per chip, but according to Intel, the effective core capacity has been increased with more flexible and efficient memory management and many neurons can be simulated with those cores.
Support for neuronal models has also been expanded. The neuron model includes a short series of microcode instructions that describe the behavior of a single neuron. Loihi 1 is optimized for a single spiking neural network but the second version allows for programmable neuron models. Neuron cores communicate with each other with spike messages and with the new version of the chip this is possible with a 32-bit spike payload that has hardware acceleration so that it doesn’t cost more power.
Intel doubled the number of processors per chip from three to six. These processors are for traditional C and Python processing and with the previous Loihi chip, the three processors often formed a bottleneck, Intel admits.
Moreover, the possibilities of communicating with the outside world are greatly expanded. Instead of a proprietary interface for communicating with other systems, Loihi 2 supports various Ethernet standards, GPIO and SPI, among others.
Intel currently provides two different Loihi 2 systems to research institutions through the Neuromorphic Research Cloud. The first is the Oheo Gulch system with a single chip and can be accessed via Arria 10 fpga and Ethernet. The second is not yet available. This is Kapoho Point, a compact system with a 4″ x 4″ form factor, eight Loihi 2 chips, an Ethernet interface, and a GPIO. Multiple Kapoho Point systems can be connected together.
In neural computing, manufacturers are focusing on systems that, like human brains, can strengthen connections in a network in order to “learn” and process data more efficiently. Loihi is a network on a chip that can be used to perform certain computational tasks faster and more economically. To simplify writing applications that can run on neural platforms, Intel provides Lava, an open source platform independent of the platform.
|Processing||14 nm||Intel 4 (pre-production)|
|Surface||60 mm²||31 mm²|
|base surface||0,41 mm²||0.21 mm²|
|transistors||2.1 billion||2.3 billion|
|the above. neuroncores per chip||128||128|
|the above. processors per chip||3||6|
|the above. neurons for each slice||128.000||1 million dollars|
|the above. Clamps for each chip||128 million||120 million|
|Go via Neuroncore||208 KB, fixed allocation||192KB, flexible customization|
|neural model||generalized LIF||Fully programmable|