Neuromorphic computing: promising innovation with tough market challenges

NC can refer to a spectrum of technologies ranging from “brain-inspired” to even traditional artificial neural networks (ANNs). However, Omdia discusses NC in terms of “biologically plausible” compute technologies rather than ANNs. Specifically, we assess the use of event-based timing, especially via spiking neural networks (SNNs) at the hardware level (i.e., asynchronous timing rather than clock-based timing).
 
Commercially, the primary benefit of NC technology is greater energy efficiency, with gains typically estimated at orders of magnitude improvement over non-NC solutions. However, much of this benefit is possible without fully event-driven hardware. Adopters also face a significant burden when using asynchronous solutions.

Other NC technology benefits include lower latency, reduced model size (and therefore die area), increased accuracy, and incremental online learning. However, each is context-specific, and wide applicability may be lacking. Adopters should investigate whether these other benefits apply to their specific context, but NC vendors may not benefit from the widespread adoption of these capabilities.
 
Considering only NC chip vendors there are only a dozen vendors that fit the NC definition as used here. Apart from IBM and Intel, these vendors are small, private organizations. Omdia notes that several large companies that had announced NC projects in the 2010s have discontinued this work or at least no longer publicly mention these efforts.
 
The NC vendors considered here are mostly focused on the edge AI opportunity, especially on the “extreme edge” share of this market. These are often Internet of Things (IoT) devices but also include smartphones and augmented reality/virtual reality (AR/VR) device classes. These vendors typically do NOT focus on the cloud and data center market.
 
NC faces key obstacles in the cloud/data center, including entrenched incumbents (i.e., NVIDIA GPUs and hyperscaler custom ASICs), adopter challenges in working with asynchronous hardware, and the ability of developers to gain much of the energy efficiency benefit of NC without using event-based processing.

Omdia estimates that the number of processors shipped each year to power devices addressable by NC technology will rise from 3.336 billion units globally in 2023 to 6.223 billion units in 2028. However, given current market features and likely trends, it is unlikely NC technology (as defined here) will account for more than a very small share of this huge amount.

To read more insights and analysis covering market trends and industry forecasts prepared by Omdia’s AI analysts, click here.