Deep Vision, a recently launched artificial intelligence startup focused on developing an AI inferencing chip for edge computing applications, is publicly announcing its emergence today. The company, established six years ago, introduces its ARA-1 processors, engineered to deliver an optimal combination of minimal latency, power efficiency, and processing capability suitable for a wide range of devices, from sensors and cameras to complete edge server systems.

Leveraging its expertise in real-time video analytics, the company is targeting its chip towards applications in intelligent retail environments, including automated stores, smart city initiatives, and Industry 4.0/robotics solutions. Deep Vision is also collaborating with suppliers within the automotive sector, with a focus on monitoring driver attentiveness – ensuring focus on the road and detecting signs of distraction or drowsiness – rather than fully autonomous driving systems.

The company’s origins lie with its CTO, Rehan Hameed, and Chief Architect, Wajahat Qadeer​, who later brought on Ravi Annavajjhala, formerly of Intel and SanDisk, as CEO. Hameed and Qadeer initially conceived Deep Vision’s core architecture during their doctoral research at Stanford University.

“They created a highly effective architecture for AI that significantly reduces data transfer within the chip itself,” Annavajjhala stated. “This results in exceptional efficiency – both in terms of performance relative to cost and performance per watt – when handling AI workloads.”

Prior to developing functional hardware, the company prioritized the creation of its compiler to confirm its ability to meet the specific requirements of its customer base. The chip design was finalized only after the compiler’s capabilities were validated.

As Hameed explained, Deep Vision’s primary objective has consistently been minimizing latency. While competitors often prioritize throughput, the team believes that latency is the more critical factor for edge computing solutions. Architectures designed for throughput are well-suited for data centers, but Hameed contends they may not be ideal for edge applications.

“[Throughput-focused architectures] necessitate processing a large volume of data streams concurrently to maximize hardware utilization, typically through batching or pipelined execution,” he clarified. “This is the only way to achieve high throughput. However, the consequence is increased latency for individual tasks, making them, in our view, less suitable for edge use cases where real-time responsiveness is essential.”

To achieve this level of performance – and Deep Vision asserts its processor exhibits significantly lower latency compared to Google’s Edge TPUs and Movidius’ MyriadX, for instance – the team employs an architecture that minimizes data movement within the chip. Furthermore, its software optimizes the overall data flow within the architecture based on the specific task at hand.

“In our design, rather than embedding a specific acceleration strategy directly into the hardware, we have incorporated the necessary programmable elements into our processor, enabling the software to efficiently map any data flow or execution flow found within a neural network graph onto a consistent set of fundamental elements,” Hameed explained.

This allows the compiler to analyze the model and determine the optimal mapping onto the hardware, maximizing data flow and minimizing data movement. Consequently, the processor and compiler can support a wide range of neural network frameworks and optimize their models without requiring developers to account for the hardware limitations often encountered with other chips.

“Every component of our hardware/software system has been designed with two overarching goals in mind,” Hameed said. “The first is to minimize data movement to enhance efficiency. The second is to maintain flexibility throughout the design, allowing for the application of the most appropriate execution plan for any given problem.”

Since its inception, the company has secured approximately $19 million in funding and filed nine patent applications. The new chip has been available for sampling for some time, and despite already serving a few customers, the company chose to remain private until now. Deep Vision anticipates that its distinctive architecture will provide a competitive advantage in a rapidly evolving market. In addition to established players like Intel’s Movidius chips (and custom chips from Google and AWS for their respective cloud platforms), numerous startups are also competing in this space, including Hailo, which recently launched its new chips after raising a $60 million Series B round earlier this year.