REPORT: Developer Spotlight – Cephable
Table of Contents
Developer Spotlight: Cephable
Cephable is an accessibility software platform that allows users to control technology using various input methods beyond the traditional keyboard and mouse. Cephable’s software enables users to control computers, applications, and digital devices using alternative inputs like voice commands, facial expressions, head motions, body gestures, and virtual buttons. Their software makes computing accessible to customers by providing adaptive voice controls, dictation capabilities, and the ability to map different physical inputs to specific actions or commands.
To date, Cephable has used CPU and GPU resources available to them on both Windows and Mac and have noticed some constraints around system performance and latency in specific use cases. By leveraging the Hexagon NPU built onto the SoC in Snapdragon X Elite they found 15% less CPU utilization and 20% less GPU utilization than an x86 system running an Nvidia 3070 ti. Compared to the M3 Macbook Pro, Cephable found 30% less CPU utilization and 50% less GPU utilization in their average user scenario workload testing. 1Cephable “average real-world scenario” which is the user running: Powerpoint (with video and images in it) Visual Studio Code (with the same 10 tabs open) Microsoft Edge (same 20 tabs open) Cephable app running with Camera Controls actively running.
Regarding choosing to utilize the Hexagon NPU on the Qualcomm Snapdragon X Elite system, CEO of Cephable Alex Dunn said:
“We’ve been developing accessibility and assistive software for years now. When we learned where Copilot+ PCs were going with including NPUs, we knew it would present an opportunity to unlock even more experiences for our users. As we’ve moved more of our offline machine learning capabilities in the Cephable app like voice, face, and head controls to the NPU, we’ve seen a substantial drop in the CPU and GPU usage which allows for more complex applications to be used, better graphics settings in games, and overall better performance for our users. We’re seeing the potential to empower our users to do more, faster, and for longer from daily digital tasks, productivity, gaming, and beyond.”
In talking with Alex Dunn about the integration of on-device AI and machine learning via a lot of computer vision, it was interesting to find out Cephable often leverages persistent inference in the background. The software uses cameras and microphones locally on device and analyzes customer face and head gestures and speech audio in real time for user interface input mechanisms. By offloading much of this inference workload to the Snapdragon X Elite NPU, Cephable found increased performance with lower latency, improved accuracy, and much lower power consumption allowing the customers to use their software for longer periods of time on battery. Offloading this workload to the NPU also allowed for a much lower CPU and GPU consumption which meant users could use more apps with more complicated workloads without compromising.
This use case solidifies that using the NPU to free up CPU and GPU resources in AI heavy applications enables a much better experience for a host of applications since it lets the CPU and GPU do their jobs more efficiently as well. It also highlights an area where the NPU is uniquely suited, which is sustained inference for long periods of time. Due to the significant power efficiency advantage the NPU has to execute AI operations at extremely low-wattage, any AI task that will run for long periods of time, either in the background doing inference, or on a workload that runs for longer periods of time, the NPU is perfectly designed for to handle that task.
Given all we have heard about NPUs through the years, we were excited to have the chance to test NPU workloads more specifically and see if they did have any real-world workload advantage in areas like speed, performance, and energy usage. The focus of our testing is always to bias toward wattage and energy efficiency and we took a similar approach when benchmarking the Snapdragon X Elite NPU.
Next ❯ The NPU Wattage Advantage