The AI accelerator Spinnaker – short for Spiking Neural Network Accelerator – has a famous father: ARM co-founder Steve Furber is behind the accelerator. What as scientific project as part of the Human Brain Project started in Manchester, Spinncloud from Dresden will soon be selling. Up to date International Supercomputing Conference (ISC-HPC), the start-up shows the second, further developed version of the chip.
It differs significantly from the established AI accelerators: Spinnaker 2 is not derived from graphics cards; each chip consists of 152 ARM M4F computing cores with self-designed accelerators for matrix operations as well as exponential and logarithm functions. They clock at a maximum of 500 MHz. The way Spinnaker works is also significantly different from other AI chips. Calculations here are not carried out using large matrices, but are event-based.
Only when the output of a simulated neuron changes will other neurons that are linked to it also be recalculated. This principle is inspired by the way the brain works; Spinnaker is one of the so-called neuromorphic accelerators. First of all, it should ensure higher energy efficiency, as the clock and operating voltage can be regulated per core. Only when an event occurs does the corresponding core increase its clock speed – no GPU can do this with such fine granularity.
Models can develop in the company
In addition, Spinnaker 2 should also be able to implement neural networks that can hardly be implemented with classic CPUs and GPUs. Spinncloud particularly emphasizes the dynamic further development of a model. This would allow new knowledge to be permanently integrated.
The advantages of expert systems and neural networks should also be able to be combined using so-called non-axiomatic logic. Expert systems contain a fixed, unchanging set of knowledge that could help against hallucinations.
Scalable from robots to supercomputers
Even if 152 computing cores sound like a lot, a single Spinnaker 2 chip is not particularly powerful. And that is entirely intentional, as the chips are intended for a wide range of uses. Combinations of one to thousands of chips are possible. In order to compete with AI supercomputers with the relatively small chips, Spinncloud soldered 48 of them onto a circuit board. Over 1,000 of these can be combined to form a supercomputer, with up to 10 million cores promises Spinncloud.
Spinnaker 2 should be available from the second half of the year, a first System with 5 million ARM cores is already installed at TU Dresden. The chip was also developed here. Further systems are to be installed in Munich, Göttingen and at the Sandia National Laboratory in the USA. The chips are also presumably manufactured in Dresden – they are created in Globalfoundries’ 22 nm FDX process.