Neuromorphic technologies will meet AI workload demands

The three main neuromorphic market segments will be consumer, industrial and automotive says Yole’s Neuromorphic Computing and Sensing Report 2021..

Until 2025, neuromorphic for industrial applications will remain a niche market with US $2 billion in 2030 for computing and sensing combined. 

Mobile and other consumer applications will reach $2.8 billion in 2030. 

 Neuromorphic computing for automotive will reach $2 billion in 2030.


Neuromorphic technologies will meet AI workload demands

Nowadays, there is a strong need for Power-efficient technologies to handle in a sustainable manner demanding AI workloads. Neuromorphic technologies are a promising answer to this need as they can perform challenging AI tasks very efficiently. 

The neuromorphic ecosystem consists today of three main categories of players: university & research institute, labs affiliated to large companies and start-ups.

Intel and IBM  have developed neuromorphic communities around their chips to help the software ecosystem grow. 

 AI is hungry for performance, and Moore’s law dynamics will not cover the needs of the ongoing 5G/IoT /AR /Robotics revolution. 

“Brute force is currently being used to leverage the power of AI, but this approach is not scalable – it will hit a heat wall, a data wall, and a cost wall,” says Yole’s AdriennSanchez. 

Neuromorphic computing and sensing solutions, mimicking the brain, have key specificities to compete within the existing AI landscape and constraints. 

These technologies will address most of the current challenges and could represent 20% of all AI computing & sensing by 2035. 

Industrial applications will be the first to use neuromorphic technologies, driven by high speed, low latency, and offline learning enabling more autonomous features and performance. 

Players such as Prophesee, Brainchip, and Nepes AI/ General Vision already have products in the market targeting industrial applications, and more players will follow in the coming years. 

The consumer market will also benefit from neuromorphic technologies enabling more AI applications at the edge on battery-powered devices, ensuring privacy and safety of personal data.

 “Current neuromorphic device architectures can also vary significantly with respect to the organization of the memory and computing components on the silicon chips,” says Yole’s Simone Bertolazzi, “there is currently a clear trend towards “in-memory-computing” solutions: several companies are developing designs with mainstream embedded memory, like SRAM, distributed across cores or neurons; various players are also considering the adoption of emerging NVM elements assembled in crossbar arrays, leveraging the “synapsis-like” properties of resistive memories, for example PCM, OxRAM, CBRAM”. 

“In the automotive market, a host of applications will benefit from the low latency and low power consumption of neuromorphic technologies,” says Yole’s Pierre Cambou, “while it will take longer for neuromorphics to be adopted in this promising yet challenging market, some projects have already been announced, such as Xperi’s Driver Monitoring System and Terranet’s ADAS cameras and laser.” 

Additionally, the cloud server market could also benefit from neuromorphic computing technologies, leveraging low latency and online learning to improve the performance of applications such as cybersecurity and fraud detection. 

The considerable power efficiency could also help to limit the power consumption growth in data centres which is a growing concern. 

Large players such as Intel and IBM are already creating neuromorphic server prototypes by assembling their massively scalable Loihi and TrueNorth chips, respectively. 

Start-ups are the first players to bring products to the market for edge computing, targeting industrial, automotive, and consumer applications. They will test new approaches in the market in a real-life environment. 

Universities are forming extensive collaborations, often supported by governments, to develop the technology and understand the hard-science potential. 

This covers a vast range, starting from a simulation of the brain on a silicon chip, and involves partnerships with various companies to develop proofs-of-concept in the field. 

Labs affiliated with large companies are deeply involved in these collaborations and often take the lead. Intel and IBM developed neuromorphic communities around their chips to help the software ecosystem grow, increase the maturity of neuromorphic AI, and test use cases directly with application players.