This market research report was originally published at JP Data’s website. It is reprinted here with the permission of JP Data.
The AI chip market continues to grow. Driven by always-on applications such as keyword detection, a whole new market segment for AI chips that consume less than 100 milli watt power has emerged in the past two years. These chips have already shipped in multi-million volume and the market shows no sign of slowing down.
Top AI chip companies such as Nvidia, Intel and Google are absent in this market segment and start-ups have capitalized on this opportunity. Several AI chips from start-ups are in production and their projections look rosy. California based Syntiant, for example, raised $55 million recently in Series B, suggesting that there’s much more growth to expect down the road.
The market is driven primarily by speech and sensor applications today. Speech applications include keyword recognition such as Alexa, background noise elimination and command recognition. Sensor driven application include activity detection, fall detection or anomaly detection. Selected vision applications such as occupancy and person detection are computationally feasible at this power envelope and vendors are actively trying to enable more vision applications. Wearables, hearables and mobile are some of the devices incorporating these chips today.
CPU and DSP based architectures are the most popular choice at this power envelop. Several applications requiring ultra-low power chips are still based on classic ML techniques that don’t necessarily need the high compute requited by neural network. The market is also rapidly evolving so the flexibility provided by CPU and DSP suites better at this power envelope. CPUs and DSPs also are useful for pre and post processing of application pipeline.
JP Data has recently published a new market report on ultra-low power AI chips. According to the report, and the market for ultra-low power chips is expected to increase at a CAGR of 65.5% from $161 million in 2021 to $3.3 billion by 2027. This CAGR is the highest of all AI chip market segments (data center/mid-power/low-power).
We expect Neural Network (NN) usage to increase in the ultra-low power chip world in the future. To facilitate, the industry will require robust software development flow which remains fragmented today. NNs will also require highest possible performance within the available power envelope. We expect this to lead to innovative chip architectures to provide needed accuracy, memory footprint and performance in the segment.
Not all is rosy for the ultra-low power AI chip market however. Vendors are likely to face competition from low-power AI chip vendors which include big names such as Qualcomm and Ambarella. The application processors in 1W power envelope offered by these companies include ultra-low power mode that offers similar functionality. Device manufactures might choose not to use additional ultra-low power chip to on save bill of materials (BOM) and could just stick with the application processor. Ultra-low power chip vendors are betting on the fact that their specialized nature will lead to new use cases and capabilities that will add value to these devices. Only time will tell which way the market will go.
The AI chip market continues to grow while enabling new and exciting applications, and ultra-low power chips add one more dimension to that. Here’s to a new class of chips!
Anand Joshi
Independent Consultant and Managing Director, JP Data
