Microchip Technology Inc., through its Silicon Storage Technology (SST) subsidiary, announced that its SuperFlash memBrain Neuromorphic Memory solution has addressed this issue for WITINMEM neuroprocessing soCs, the first mass-produced sub-MA system, Can reduce speech noise and recognize hundreds of command words. In real time and immediately after power-on. In-memory computing techniques are balanced to eliminate a large number of data communication bottlenecks that would otherwise be associated with performing AI voice processing at the edge of the network. However, it requires an embedded memory solution that simultaneously performs neural network calculations and stores weights.
The company partnered with WITINMEM to incorporate its memBrain analog memory computing solution based on SuperFlash technology into its ultra-low power soCs. The SoC has in-memory computing technology for neural network processing, including speech recognition, deep speech denoising, voice print recognition, scene detection and health status monitoring. In turn, WITINMEM is working with multiple customers to bring products based on this SoC to market by 2022.
"WITINMEM's memBrain solution is a groundbreaking work that addresses the computation-intensive needs of real-time AI speech at the edge of the network, based on advanced neural network models," said Shaodi Wang, CEO of WITINMEM. "We pioneered the development of an in-memory computing chip for audio in 2019, and now we have achieved another milestone by mass-producing the technology in our ultra-low-power neural processing soCs, simplifying and improving speech processing performance in smart speech and wellness products."
"We are pleased to have WITINMEM as a major customer and applaud the company for using our technology to enter the expanding AI edge processing market with a superior product," said Mark Reiten, VICE President, LICENSING, SST. "WITINMEM SoC demonstrates the value of using memBrain technology to create a single-chip solution for memory-based computational neural processors that eliminates the problematic machine learning model of traditional processors that use digital DSP and SRAM/DRAM based methods for storage and execution."
Microchip's memBrain Neuromorphic memory product is optimized for VMM for neural networks. It allows battery-powered and deeply embedded edge processors to deliver the highest AI reasoning performance per watt. This is achieved by storing the neural model weights in an in-memory array and using the in-memory array as neural computation elements. Because external DRAM and NOR are not required, the result is 10 to 20 times lower power consumption than other methods and lower overall processor BOM costs.
The neural models permanently stored in the processing elements of the solution support the just-in-time capabilities of real-time neural network processing. WITINMEM uses SuperFlash technology's non-volatile floating gate unit to close its in-memory computing macros by idle state to further reduce leakage power in demanding iot use cases.
