AI is disrupting almost every sector in the digital economy, from agriculture to retail, manufacturing and mobility, but its deployment is challenged by high hardware costs and software complexity. NeuReality, an Israeli startup, is working on a solution to this problem by designing a new system architecture for AI systems. This radically different system architecture can drastically reduce their cost and energy consumption.

During the inference stage of AI deployment, data processing requirements are unique. The industry has invested heavily in developing optimized AI chips, but they operate in architectures that are heavily dependent on CPU and networking chips. A system board with - say - $4,000 worth of AI chips needs three times that amount in processor and networking chips to support them, with the CPU often being the bottleneck, limiting the output of the AI-specialized chips. NeuReality’s technology is based on a completely new AI-centric system architecture designed to be optimized for AI processing, replacing the traditional CPU-centric support system.

The company plans to reduce the price tag for the chips needed to support the AI chip from $12,000 to $1,400, and improve the support for the AI chips, resulting in better performance. In addition to the hardware savings, the overall system will have a much reduced energy footprint and data center footprint, resulting in further cost savings. The company also addresses the software complexity barrier with a software development kit that will make it much easier for data scientists and DevOps engineers to deploy AI.

The market for AI chipsets is large and growing rapidly. Market survey company Omdia estimates that by 2025 the total addressable market will reach $75 billion, with applications in cloud and data centers, near edge, and edge applications.

NeuReality has partnership agreements with Lenovo, IBM Cloud, Xilinx and others. The company forecasts initial sales of $2.5M in 2022 based on its prototype released in 2021, growing to $30M in 2024.

Photo credit: Markus Spiske