Brain-inspired chips are about to usher in the post-Nvidia era
Neuromorphic chips consume 1,000 to 15,000 times less power than Gpu chips. And from the labs they are reaching the market
3' min read
3' min read
Generative artificial intelligence has a problem: it consumes too much energy. Despite successes in imitating human reasoning, on the energy efficiency front, AI has so far lost the challenge to the brain.
Nvidia's powerful Gpu chips have made it possible to build supercomputers for parallel computing capable of training large language models. The technology that, from ChatGpt onwards, gave machines the ability to understand language and generate text and images autonomously. Like all innovations, however, these chips are destined to be superseded. Among the possible replacements are neuromorphic chips, circuits designed to mimic the functioning of the smartest machine of all that consumes just 20 watts to operate: the human brain.
Model inspired by the brain
.If you can't beat it, imitate it. This can be summarised as the race for computing architectures that will open the post-Nvidia era and that are inspired by the human brain. Animal neurons do not activate all at once, but individually and only in the presence of a stimulus. In the neural networks in use today, by contrast, all the synthetic neurons in a Gpu are always active. Hence the idea behind neuromorphic computing: imitate animal behaviour and only activate the computing nodes when the situation calls for it. Translated into hardware, this means not using the whole network to do the calculation, but only a part. Or even, following a second line of research, enclosing both the memory unit and the calculation unit on the chip. Just like in neurons.
Less energy consumption
.Vittorio Fra is an engineer and researcher at the Electronic design automation Group of the Politecnico di Torino. He is one of the few Italians who spoke at the latest Neuro inspired computational elements conference (Nice), the 13th annual conference dedicated to neuromorphic computing held at the University of Heidelberg, Germany. Researchers from all over the world announced novelties, shared results, presented applications: concrete use cases, run on both anthropomorphic and traditional hardware to compare performance.
'If one measures,' Fra explains, 'the power consumed to run an artificial intelligence algorithm on neuromorphic chips and compares it with that required to run the same algorithm on traditional Gpu-based hardware, it is easy to perceive what a great advantage neuromorphic computing has. The savings depend on the type of calculation but, all things being equal, the energy consumed is a thousand to 15 thousand times lower. All this without affecting either the speed of calculation or the correctness of reasoning'. The Turin group is part of the Ebrains-Italy network and has the task of developing an infrastructure for prototyping neuromorphic solutions.
