Edge AI in a package for consumer designs
AI accelerator developer Femtosense has launched a system in package (SiP) for consumer applications.
The AI-ADAM-100 SIP combines the Femtosense Sparse Processing Unit 001 (SPU-001), a neural processing unit (NPU), and an ARM Cortex M0+ microcontroller from Abov Semiconductor to provide edge AI voice processing and voice-cleanup capabilities.
This follows a deal signed between Femtosense and Abov at the CES show in January. “One of the reasons we can do this so quickly is we can combine the SPU001 die with MCU in a SiP package,” Sam Fok, CEO of Femtosense tells eeNews Europe.
Partnership brings energy saving AI technology to home appliances
The Abov 40nm M0+ microcontroller has 8Kbytes of SRAM and 32Kbytes embedded flash as the controller. It can use the 1Mbit of SRAM in the 22nm SPU-001 as overflow, for example for recording a second of voice for an always on system.
This is the first in a series of SiP devices using microcontrollers with ARM M4 and M7 microcontrollers. “Cost is the primary driver, more than power and definitely more than form factor and its targeting consumer equipment and we are looking at 250uW to 5mW so its really the cost that is the issue,” says Fok.
The SIP uses an organic substrate and is assembled by Winpak, in which Abov recently took a stake. “Winpak did commodity memory packaging and test and so this was a three way deal,” he said. “We provide tested wafers to the Winpak consortium.”
The AI-ADAM-100 MCU provides edge AI voice processing and voice-cleanup capabilities on-device using the SPU-001 AI accelerator. This uses a custom instruction set architecture with large vector ALUs to handle matrix multiplication in parallel.
“That allows us to save on memory operations and we have custom memory formats so you don’t store data you don’t have to,” Fok tells eeNews Europe. The weights and activations in the AI framework can be compressed to preserve the performance of the large network
“We also work on compressed data directly. INT4 is the lowest precision and because we are in the audio space we have activations with higher precision, so we have INT16 activations with INT8 weights, INT8 activations with INT8 weights and INT8 activations and INT 4 weights, so different applications can use different weight levels.”
Sparse AI reduces the cost of AI inferencing by zeroing-out irrelevant portions of an algorithm and then only allocating hardware memory and compute resources to the remaining nonzero, relevant portions of the algorithm.
The AI-ADAM-100 allows appliances can implement “say what you mean” voice interfaces that allow users to speak naturally and express the intent freely in multiple ways. For example, “Turn the lights off”, “Turn off the lights,” and “Lights off” all convey the same intent and are understood as such. Voice/audio cleanup processes data before it is sent to the cloud, improving reliability and accuracy while reducing the volume of data sent, thus reducing backend infrastructure costs.
“This is targeted to Abov’s main market for smart devices, we can do wake up word. Having more models in a smaller form factors means you can do more functions, for example the same command in multiple ways. For dual mode capability we can also work on audio clean up from multiple microphones and compress to save the processing costs in the cloud.”
“With sparsity integrated throughout the AI development stack, the AI-ADAM-100 is the first device on the market to fully unlock the advantages of sparse AI,” he said. “Our sparsity-enabling technology allows our customers to deliver compact, efficient AI processing to a growing variety of markets and products, including home appliances as well as small form factor, battery-operated devices like high-fidelity hearing aids, industrial headsets, and consumer earbuds.”
On top of the AI-ADAM-100, Femtosense provides software for tool-driven edge AI applications or full custom implementations using a manufacturer’s own AI models, whether dense or sparse.
“The AI-ADAM-100 is the best-optimized AI MCU solution for voice and audio-based AI applications and enables a variety of on-device AI applications for consumer electronics and standalone devices,” said Choi Won, CEO of ABOV Semiconductor. “Together with Femtosense, we will continue to develop the most cost- and power-efficient AI MCUs for global customers.”
ABOV has verified AI-ADAM-100’s voice command recognition performance under multiple noise conditions, meeting the requirements of leading customers. Global home appliance makers are working to reduce the number of buttons on devices and streamline the user experience using AI-based voice commands.
Engineering samples of the AI-ADAM-100 are available now with commercial mass production targeted for later this year. Development support includes software tools, evaluation boards, and demo AI models, including Smart Home Appliance Wake-up and Control.
femtosense.ai/products/; www.abov.co.kr.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
