AI started taking over the ultraportable world, with current “system-on-a-chip” units allowing for machine learning tasks to be done without an internet connection. Still, these chips are expensive and only available to high-end devices. That will soon change, as Femtosense’s newest AI-ADAM-100 system-in-package neural processing unit allows efficient AI voice processing.
AI-ADAM-100 is designed in collaboration with ABOV Semiconductor and combines with an Arm Cortex M0+ microcontroller and a Sparse Processing Unit (SPU). SPUs use sparsity techniques to optimize machine learning models for devices with limited computing power, such as edge devices (smartphones, IoT devices). Furthermore, the system-in-package solution has 64/32 KB MCU code flash memory, 8 KB SRAM, and Femtosense’s SPU-001 NPU with 1 MB of NPU SRAM. The NPU is built on a 22-nm process, while the MCU is built on a 130-nm process.
What is AI sparsity you might ask? Well, many parts of a model’s data (matrices or tensors) contain values that are zero or insignificant. When these values are zero, they don’t need to be calculated, allowing the device to skip certain computations. This dramatically reduces the amount of processing required, making it more efficient for smaller devices to run AI models without needing extensive power. In addition, sparse weight matrices can be compressed in memory. Thus, while Femtosense’s NPU has 1 MB of SRAM, the effective value with sparsity is 10 MB. In other words, these chips have excellent value-to-performance ratio and consume way less electricity.
Femtosense sees its latest innovation as part of various devices. For instance, it can be built into hearing aids, industrial headsets, consumer earbuds, and even home appliances. Utilizing this chip’s abilities, the user can talk naturally to the device, but the chip can also analyze the surrounding noise and provide noise cancelation.
Femtosense plans mass production of the AI-ADAM-100 system-in-package unit later this year. However, developers and manufacturers can acquire engineering samples directly from the company right now.
PR_2024-07-17.pdf (femtosense.ai)
