Computing-in-memory technology is addressing the vast data communications bottlenecks that are associated with performing AI speech processing at the network’s edge. It requires an embedded memory solution that can run neural network computations and store weights. Silicon Storage Technology, a Microchip Technology subsidiary announced SuperFlash, memBrain. The control system is a neuromorphic memory solution that has solved the AI speech processing problem for the WITINMEM neural processing SoC. This is the first volume production that enables sub-mA systems to reduce speech noise and allow for recognition of hundreds of command words in real-time and immediately after start-up.
memBrain is an analog solution incorporated into WITINMEM’s ultra-low-power System on Chip. The Soc is built to run computing-in-memory technology for neural networks processing, for such applications as speech recognition, voiceprint recognition, deep speech noise reduction, health status monitoring, and scene detection.
Optimized for performing vector-matrix multiplication (VMM) for neural networks, memBrain, is now the go-to solution for WITINMEM, which is already volume producing the SoC. The system enables processors used mainly in deeply-embedded edge devices that are battery-powered to run the highest possible AI inference performed per watt. The way this is accomplished is by storing the neural model weights as values in the memory array. The memory array is then used as the neural compute element.
The results in power consumption are 10 to 20 times less power than any other approach alternative. The process also lowers the overall processor Bill of materials (BOM) costs, because DRAM and NOR are unnecessary. The neural models are permanently stored in the processing element of memBrain. This supports instant-on functionality in real-time for neural network processing.
This technology will vastly change how AI functions in real-world, everyday, walk-of-life situations, such as more virtual assistants with greater width and breadth of recognition of the user. There can be further recognition of the user, and multiple related users, in handheld devices. In the medical setting people with a need to access records in time-constrained situations, such as during an emergency, can be recognized in real-time through voice patterns and able to initiate life-saving processes hands-free.
Microchips technology’s SST subsidiary is a provider of embedded flash technology. The company offers a range of SuperFlash memory technology solutions for consumer, industrial, automotive, and IoT markets.
For the past 8 years, I have been writing about embedded technologies, added to my technical, academic, and medical editorial experience, with companies like Elsevier and Cambridge University Press. I tell people to read what I write, not try to pronounce my last name. I am always available for comments and suggestions you can reach me at email@example.com and I promise I will take the time to reach back out to you. I live in the North East with my wonderful family.