Machine learning (ML) is all about massive amounts of processing, DSP, etc., right? Maybe not, according to the team at Aspinity. The company continues to push ahead on the analog front. The latest member of the company’s analogML family, the AML100, operates completely in the analog domain. As a result, it can reduce always-on system power by 95% (for the record, we had to walk through this a couple of times before I believed them).
The AML100 reduces always-on system power to under 100 μA. Such a power level makes the device suitable for battery-powered ML applications. That could be security devices that are affixed to a wall, biomedical monitoring, or voice-enabled systems.
Minimizing the quantity and movement of data through a system is a very efficient way to reduce power consumption, which is the concept behind the analog ML. And the data coming in is already in an analog form, so you are eliminating a step.
The AML100 delivers a substantial system-level power-savings by moving the ML workload to low-power analog, where the device can determine data relevancy with a high degree of accuracy and at near-zero power. This makes the AML100 the only tinyML chip that intelligently reduces data at the sensor while the data is still analog and keeps the digital components in low power mode until important data is detected, thereby eliminating the power penalty of digitization, digital processing, and transmission of irrelevant data.
The heart of the AML100 is an array of independent, configurable analog blocks (CABs) that are fully programmable within software to support a wide range of functions, including sensor interfacing and ML. The AML100 is highly flexible, and can be reprogrammed in the field with software updates or with new algorithms targeting other always-on applications.
The precise programmability of the AML100’s analog circuits also eliminates the chip-to-chip performance inconsistencies typical of standard analog CMOS process variation, which has severely limited the use of highly sophisticated analog chips, even when the inherent low power of analog makes it better suited for a specific task.
The AML100, housed in a 7- by 7-mm 48-pin QFN package, is currently sampling. Volume production is planned by the end of the year. Integrated hardware-software evaluation kits are available: the EVK1 for glass break and T3/T4 alarm tone detection or the EVK2 for voice detection with pre-roll collection and delivery.
http://mpvideo.qpic.cn/0bc3iqaaaaaai4ahkotcjnrfargdabcaaaaa.f10002.mp4?dis_k=6ab1141b46808b0f52a54b2aaf8cf9cb&dis_t=1652764128&vid=wxv_2366046820755374082&format_id=10002&support_redirect=0&mmversion=false
