Neuromorphics will boost computional efficiency according to Yole
“Many similarities point to the idea that such a paradigm shift could happen quickly” Cambou adds.
Several years ago, the biggest obstacle preventing the Deep Neural Network (DNN) approach from performing its best was the lack of suitable hardware to support DNN’s innovative software advances. Today, the same is true for neuromorphic technology – but as the first Spiking Neural Network (SNN) chips roll out, the first beachhead markets are ready to fuel growth. The initial markets are industrial and mobile, mainly for robotic revolution and real-time perception.
Within the next decade, the availability of hybrid in-memory computing chips should unlock the automotive market, which is desperate for a mass-market AD technology.
Neuromorphic sensing and computing could be the magic bullet for these markets, solving most of AI’s current issues while opening new perspectives in the decades to come, says Yole.
In its new report, Neuromorphic Sensing & Computing, the market research firm explores the computing and deep learning world with an imaging focus, delivering an in-depth understanding of the neuromorphic landscape with key technology trends, competitive landscape, market dynamics and segmentation per application.
Since 2012, deep learning techniques have proven their superiority in the AI space. These techniques have spurred a giant leap in performance, and have been widely adopted by the industry.
“Recently, we have witnessed a race for the development of new chips specialized for deep-learning training and inference, either for high-performance computing, servers, or edge applications”, asserts Yohann Tschudi, Technology & Market Analyst, Computing & Software at Yole. “These chips use the existing semiconductor paradigm based on Moore’s Law. And while it is technically possible to manufacture chips capable of performing hundreds of Tops to serve today’s AI application space, the desired computing power is still well below expectations.”
Consequently, an arms race is ongoing, centring on the use of ‘’brute force computing” to address computing power requirements. The technology node currently used is already at 7nm, and full wafer chips have emerged. Room for improvement appears small, and relying solely on the Moore’s Law paradigm is creating several uncertainties.
Current deep-learning techniques and associated hardware face three main hurdles: first, the economics of Moore’s Law make it very difficult for a start-up to compete in the AI space and therefore is limiting competition. Second, data overflow makes current memory technologies a limiting factor. And third, the exponential increase in computing power requirements has created a “heat wall” for each application.
Meanwhile, the market is demanding more performance for real-time speech recognition and translation, real-time video understanding, and real-time perception for robots and cars, and there are hundreds of other applications asking for more intelligence that combines sensing and computing.
Given these significant hurdles, the time is ripe for disruption: a new technology paradigm in which start-ups can differentiate themselves, and which could utilize the benefits derived from emerging memory technologies and drastically improve data, bandwidth, and power efficiencies. Many foresee this new paradigm to be the neuromorphic approach, some would call it the event-based approach where computation happens only if needed instead of being done at each clock step. This method allows a tremendous energy saving essential to run these greedy and intensive AI algorithms. Yole sees this as the most probable next step in AI technology. Its most recent report represents a window into a possible future where AI uses neuromorphic approaches for sensing and computing.
Yole Développement – www.yole.fr